1 of 52

Geo for Good 2021 Poster Gallery

2 of 52

Abstract

Agricultural Land Cover

The availability of satellite imagery and Artificial Intelligence has been recognized by Statistics Indonesia as one of effective data collection tools and methodology. In Statistical Mapping Division, we develop and implement methods using satellite imagery Sentinel-2 and machine learning algorithms, CatBoost (gradient boosting tree algorithms), for identifying the potential agricultural lands in order to support Census of Agriculture 2023. The algorithm able to classify the land cover not only the national level but with enhanced within country disaggregation.

Goals

  • Developing agricultural land cover utilizing satellite imageries and machine learning
  • Providing spatial statistical frame for Census of Agriculture 2023 in Indonesia

Methods

Result

Conclusion

Lessons Learned

The output, the model able to acquire the average accuracy over 93.6% using F-1 score as evaluation metrics. However, the accuracy below 80% occurred in particular class such as mixed plantation class due to the high heterogeneity of pixel characteristics.

Indonesian Agricultural Land Cover

for Census of Agriculture 2023 in Statistics Indonesia BPS

Achmad Fauzi Bagus Firmansyah, achm.firmansyah@gmail.com

Wida Widiastuti, nenkwida@gmail.com

Data Acquisition

Data Labelling

Model Development

Data Production

Start

Divide Indonesia into 20x20 km grid

Acquire Sentinel 2 yearly composites and topographical data using GEE

Preprocessing and vegetation indices calculation

Grouping Indonesian grid utilizing ecoregions

Choose sampling grid 5% for each ecoregion

Create Data Training through visual interpretation and additional data

Removing Outlier for each label

Split Training 80% and Testing 20% with label stratification

Recursive Feature Elimination and Hyperparameter Tuning using Optuna

3-fold cross validation for model building

Record the performance

Post Classification using window filtering and calbirating with other national ministry maps

Classifying data using developed model

Finish

Next Step

  • In 2022 we will conduct a massive field survey for collecting the in-situ training samples to obtain a better result of classification
  • Developing enhanced algorithms and revised strategies to get higher performance and accuracy

Class

F1-score

Class

F1-score

Rice Fields (1)

0.9

Wetlands (6)

0.9

Non-irrigated crop fields (2)

0.86

Shrub (7)

0.87

Palm plantations (31)

0.89

Pastures (8)

0.86

Sugarcane plantations (32)

0.86

Artificial Surfaces (9)

0.91

Tea plantations (33)

0.87

Pond (10)

0..85

Other annual plantations (34)

0.86

Water Surface (11)

0.96

Mixed plantations (4)

0.78

Others (12)

0.87

Forest (5)

0.95

Average F1-score for each class nationally

Result for Java Island

Sentinel 2

Composite 2020

Agricultural Land Cover

3 of 52

Research Background

The Yawuru People are the Traditional Owners of the land and sea Country around Broome, Western Australia. Managing ecologically and culturally sensitive Country alongside a working cattle station requires Yawuru managers to balance competing priorities to protect Country and Culture. Nyamba Buru Yawuru (NBY) have developed a nation-leading geospatial mapping program¹ to assist in negotiations and decision making.

In a novel application of the SEEA-Ecosystem Accounting², we are harnessing the accessibility and compute power of Google Earth Engine to build strategic intelligence about biophysical and cultural values across Yawuru Country. We have developed an interactive dashboard to monitor biophysical change over time for vegetation, surface water and cultural burning.

Harnessing Technology

Geospatial Intelligence for Indigenous Land Management

Anna Normyle, Bruce Doran, Dean Mathews and Julie Melbourne

anna.normyle@anu.edu.au

We acknowledge the Yawuru people, the Traditional Custodians of the lands and waters in and around Rubibi

(the town of Broome). We acknowledge and respect their continuing culture and contribution to this research.

Results

  • Dashboard to support decision-making and communication of benefits of Yawuru management.

  • We are integrating cultural values for a holistic assessment of the Yawuru living cultural landscape.

ٰ¹ Australian Human Rights Commission, 2012. Native Title Report. Sydney.

² Normyle et al., 2021. Including Indigenous Perspectives in SEEA-EA in Theory and Practice. London Group on Environmental Accounting 27th Meeting.

SEEA: System of Environmental-Economic Accounting

4 of 52

What do conservation & resource managers need?

Data to fill monitoring gaps...

Conservation Districts assist with ecological monitoring of agricultural areas in the USA. They report on the status of ecosystems embedded in mostly privately-owned lands. Easier access to remotely sensed data could improve the extent and effectiveness of their monitoring programs.

What is needed?

  • Easy-to-use and stable data analysis workflow
  • Ability to separate natural areas from farm fields
  • Ability to systematically detect changes in type of vegetation over a 5-year period at watershed scales
  • A way to quantify change in terms of % area for reporting that preserves the privacy of individuals

Bringing Earth Engine into ecosystem monitoring

Identifying relevant & feasible GEE analysis

Tracking landscape change

We developed scripts in the Earth Engine code editor using Sentinel-2 satellite imagery to:

  • Classify agricultural vs. natural vegetation based on plant phenology
  • Compare cover classes across four years to quantify gains and losses of agricultural and natural vegetation
  • Map areas that are always green (naturally vegetated or always irrigated)
  • Map areas of change (green-or-senesced) which are important for reporting and ongoing management

Transferring capability to a broader user group

Current work and next steps

  • Hosting an online User Library (beta) directing users to Apps, a tutorial, or our GitHub repository, according to level of interest
  • Continuing to partner with Conservation Districts and the State Conservation Commission to identify gaps that GEE could fill
  • Working individually with conservation professionals to provide guidance as they modify scripts to fit their needs.
  • Scoping future GEE applications to inform conservation and sustainable land and resource management

Adding Earth Engine to the conservation toolbox

Amanda T. Stahl, atstahl@wsu.edu

References

Stahl, A.T., Fremier, A.K., Heinse, L., 2021. Cloud-Based Environmental Monitoring to Streamline Remote Sensing Analysis for Biologists. BioScience. https://doi.org/10.1093/biosci/biab100.

User Library (under development): http://labs.wsu.edu/ecology/research-projects/cbem-user-library/

Developing a workflow with GEE and monitoring programs

WSU partnered with two local Conservation Districts to co-develop an analysis workflow for monitoring. We worked with Sentinel-2 image classification in Google Earth Engine. The products of this initial workflow have been used by three counties for reporting to the state. We summarized our collaborative process in a series of 5 steps that could be used in any setting.

5 of 52

Introduction

Creating a Carbon Budget

  • Cities release nearly 70% of global carbon emissions into the atmosphere
  • City based carbon budgets can help in guiding policies to minimize greenhouse gas emissions
  • Satellite data can enable estimations of urban carbon budgets using land cover data but the spatial resolution of the data is of paramount importance for an accurate accounting.

Exploring the effect of Scale

  • The animation is looping through 6 different Land Cover maps [1-6] with different resolutions, starting with 5000m to 10m cell size, using the same legend
  • Each land cover class has an unique Carbon potential, and the combination of classes is the City Carbon Footprint
  • Accurate representation of land cover is necessary to create a realistic Carbon Budget

Methodology

Mapping land cover was one of the first satellite mission and the GLASS (The Global Land Surface Satellite) CDRs (Climate Data Records) project is a good example of collating LandSat AVHRR data from 1982 to 2015, but for a city carbon budget this resolution is not sufficient.

Last month the latest Land Cover product was announced, the ESA 10m world cover, with 11 distinct classes.

Besides these two, there are many more products, each with an unique combination of resolution and classification. In this study 6 different Land Cover Classification products are compared, and assessed on usability at local level, to produce the most accurate carbon budget

Process

  • Create an Area of Interest (AoI) over the selected city
  • Clip the land cover layers by the AoI
  • Reclassify the land cover classes to a uniform key (f.e., the ESA GLC 10m legend)
  • Assign a Carbon factor by land cover (tonnes CO2/hectare/year), obtained from literature [7-11]
  • Calculate the City Carbon Footprint (in CO2 Amount as Mt (million tons)/year)

Results

  • The Carbon budget in this example is created using an large area over Nottingham (UK). The main argument to select this region is the large variation of land cover (but the comparison is valid for any other area)
  • Striking is the difference between the 30m land cover classification and the other layers; this is primarily because of the difference in source data (LandSat vs Sentinel) and classification method

Conclusion

Carbon Budgeting

  • Only recently it is possible to use satellite (remote sensed) data to create accurate Carbon Budgets. For example, 240 cells were used to create the lowest resolution map, covering 7200 sq km. In comparison, the carbon balance using the highest resolution map contains 72 million cells.
  • With the new 10m resolution map it is possible to compute Carbon budgets at a much finer city scale.

Outlook

  • Reintroducing the discontinuous urban fabric from the 100m Land Cover legend
  • Also, introducing a road layer, so that the Urban Footprint can be further refined, and broken down in subclasses (for buildings and transportation)

References

Why Scale Matters - Creating a Carbon Budget

Gijs van den Dool - gvdd@everimpact.com - Connect on LinkedIn

Ankur A. Shah - ankur@mycelium.ngo - Connect on LinkedIn

Land Cover Change by spatial resolution

ESA World Cover (10m)

Land Cover

CO2 Amount

Tree cover

(0.68)

Shrubland

(0.00)

Grassland

(0.71)

Cropland

(0.35)

Built-up

2.81

Bare

0.53

Permanent Water

0.00

Herbaceous Wetland

(0.00)

Total (over 7200 sq km)

1.59

Mt (million tons)/year

”All models are wrong, but some are useful” George E. P. Box

Nottingham Wards - Carbon Budget

Estimated CO2 Amount in Mt (million tons) / year

6 of 52

Introduction

Floods

Floods are natural phenomena that have always been occurring, in some regions vital for the functioning of ecosystems and human activities such as fishing and agriculture. However, floods are also a devastating natural hazard, causing damage, displacement and loss of life, exacerbated by human-induced land use changes and climate change.

Flood Mapping

Flood maps (forecasting, near real-time or historical) can be derived from models and/or earth observation (EO) data. We will focus on those derived from EO. Regardless of the method, they can provide valuable insights into the geographic extent and progression of a flood. There are currently many providers of EO-derived flood mapping products (too many to list here), which on its own poses a problem for many end-users (e.g. governments, NGOs, humanitarian agencies), as they cannot be expected to understand the strength and limitations of each.

Actionable information needs

However, even a perfect flood map does not provide actionable information on its own, as this requires:

  • context (e.g. “how does this compare to previous years?”)
  • impact (e.g. “how many people will be affected?”)

This is often lacking or handled in a too simplistic manner for it to provide information on which important decisions can be based.

Method

We use the HYDrologic Remote sensing Analysis for Floods (HYDRAFloods) tool to create flood maps, which utilizes Google Earth Engine [1,2]. This is developed within the SERVIR-Mekong program, a partnership between USAID and NASA. The tool is operational over parts of Southeast Asia and has also been used ad-hoc in Central America. Research projects have been carried out in several other countries (e.g. Myanmar, Ethiopia).

In order to provide context to these maps we consider flood drivers (e.g. precipitation, topography) and historical water presence, through existing sources (e.g. JRC GSW, locally available data) and/or by creating custom maps. Issues or uncertainties can arise from:

  • inaccuracies in the data
  • differences in spatial and/or temporal resolution

Impact is often handled by overlaying exposure datasets (e.g. land use, population, infrastructure), but this also introduces many uncertainties. This is related to the same topics already mentioned above, but it also often misses information on flood depths and velocity, which are critical when looking at impacts (but notoriously hard to obtain from EO data).

Results

We show that the inclusion of reference data provides more detailed information and provides context, which helps end-users to better understand the flood maps.

We have also shown that impacts can vary widely based on the used exposure data, even without considering flood depths or velocity. Using a single data source can provide a false sense of security, but the uncertainty that becomes visible when using multiple sources also poses problems, as it can make it difficult to understand for end-users.

Conclusion

Context

  • By comparing flood maps against reference data we can place a specific flood in a historical context
  • The inclusion of flood drivers helps to provide confidence to flood maps, as well as identify areas of potential shortcomings or uncertainties
  • Both help end-users to better understand the maps, as well as guide their decision making process

This requires knowledge of the area, particularly hydrological expertise, preferably at the regional or local level.

So you have a flood map...now what?

Arjen Haag, arjen.haag@deltares.nl

References

  1. Markert et al. (2020). Comparing Sentinel-1 Surface Water Mapping Algorithms and Radiometric Terrain Correction Processing in Southeast Asia Utilizing Google Earth Engine. Remote Sens. 12. doi: 10.3390/rs12152469
  2. Mayer et al. (2021). Deep learning approach for Sentinel-1 surface water mapping leveraging Google Earth Engine. ISPRS J. Photogramm. Remote Sens. 2. doi: 10.1016/j.ophoto.2021.100005

Impact

  • Impacts derived by overlaying flood maps on exposure data can produce widely different results for each dataset
  • This means such an analysis is highly uncertain and requires proper framing when reported
  • Accurate impact assessments also require more information than flood extents, such as depth and velocity, which are hard to obtain from EO data

This requires further research into best practices, new exposure datasets and EO-derived flood information.

Floods in Myanmar, late July 2018. Photo: MOI Myanmar

7 of 52

SIKU: the Indigenous Knowledge Social Network, is a powerful platform developed by Inuit to support Indigenous self-determination, safety, food security and knowledge transfer.

The SIKU app makes it easy to create posts with photos, tag

Inuktut sea ice terminology, and support knowledge and

language transfer in communities! Learn more

Method

C) Ice Watch Challenge

D) Next Up: Machine Learning and Sea Ice Safety

When the computer and the harpoon work together: Using Inuit Knowledge, machine learning, Earth Engine, and satellite radar to support sea ice safety through SIKU: The Indigenous Knowledge Social Network

  1. What Is ?

People from across Inuit Nunangat took part in our first annual Ice Watch Challenge, making 295 ice posts across 32 communities. The posts documented dangerous ice, shared Inuktut terminology, and will be used to improve the SIKU Ice Map. Learn more

B) SIKU Ice Map: Inuit Knowledge, Sea Ice Charts, Satellite Radar

Our top priority is to use machine learning to find and display ikiraasak (polynyas - open water surrounded by sea ice) located within the landfast ice, followed by other ice features important for safe travel like apputainaq (open water hidden by snow) and siqummaq (mobile cracks). This project is just beginning, and will use two main data sources: RADARSAT Constellation Mission radar imagery and observations of polynyas. Interested in learning more or contributing? Contact us!

The SIKU Ice Map (top) highlights differences between a) landfast ice / mobile ice, b) growing (thinner) ice / thicker ice, and c) rough ice / smooth ice.

The SIKU Ice Map is designed to bring together the best information to help northern communities access information about ice. It currently uses SIKU Ice Posts, ice charts from the Canadian Ice Service and radar imagery from the Sentinel-1 satellites, which are combined using Earth Engine. We are working to build more Inuit Knowledge into the SIKU Ice Map by using SIKU Ice Posts made by Inuit sea ice experts to train machine learning models! Learn more

8 of 52

9 of 52

Introduction

Taking soil moisture to the next level

Soil Moisture is a key parameter in many environmental disciplines as well as it is crucial for practical decisions in agriculture. However, recent soil moisture data either lack spatial or temporal resolution, their application requires an advanced technical background or they need an extensive computing infrastructure. In this context, I am developing a soil moisture product within the Google Earth Engine that is both high resolution and easy to use, making practical agricultural decisions possible and overcoming the aforementioned problems.

Objectives:

  • High temporal (up to 1 - 2 days) and spatial (up to 10 m) resolution soil moisture maps
  • Easy to use and no local computation infrastructure necessary
  • On demand and fast delivery for individual time period, region and resolution

Method

Results

The high resolution soil moisture product based on the Sentinel-1 satellite mission gives reasonable results compared to in situ measured data. The error between estimated and in situ soil moisture was evaluated over the Rur catchment and ranges between 2.6 % and 7.0 %, depending on the vegetation coverage. Using the Earth Engine, both high resolution products for field-scale use (10 m) as well as mid-resolution products for catchment-scales (200 m) can be computed fast and on demand.

Conclusion

The Next Steps

  • Using Earth Engine for on-demand calculation of soil moisture maps is feasible and could improve agricultural management decision, with a minimum of technical background and infrastructure.
  • The method gives excellent results for various agricultural sites, especially meadows. Nevertheless, for some vegetation coverages further improvement is necessary.
  • With the integration of SAR images with a longer wavelengths (e.g. ALOS-2, ROSE-L) as well as a remotely sensed crop type classification, the effects of vegetation can be addressed.

References

  • Balenzano, A.; Mattia, F.; Satalino, G.; Davidson, M.W.J. Dense Temporal Series of C- and L-band SAR Data for Soil Moisture Retrieval Over Agricultural Crops. IEEE J. Sel. Top. Appl. Earth Observations Remote Sensing 2011, 4, 439–450, doi:10.1109/JSTARS.2010.2052916.

High Resolution Soil Moisture Mapping with Earth Observation

David Mengen, mengendavid@gmail.com

The method is based on the alpha approximation method proposed by Balenzano et al. [1], assuming that between two consecutive Synthetic Aperture Radar (SAR) recordings, a change in backscattering signal is caused mostly by a change in soil moisture, as it has the fastest rate in temporal change compared to other surface parameters. The algorithm is embedded into an automated workflow, using the cloud-processing platform Google Earth Engine. As input data, the Sentinel-1 A/B SAR image collection is used. To reduce the time between consecutive observations, all available orbits as well as both ascending and descending acquisition modes can be used. Looking a bit ahead, the method is also developed for the upcoming L-band SAR satellite mission ROSE-L from the European Space Agency.

Step-by-step

  1. Sentinel-1 image collection
  2. Speckle Filtering
  3. Incidence Angle Normalization
  4. Detrending of Vegetation Influence
  5. Low Pass Filtering
  6. Backscattering to Soil Moisture Inversion

10 of 52

11 of 52

Introduction

Watershed prioritization is the classification of a catchment’s various sub-watersheds according to the order in which soil management and conservation steps must be taken.

Need of proper soil and water management treatments to overcome continuous soil loss by runoff and other ecological problems in Kosi watershed because of lack of capability to withstand it. Also, speedy land use modification as it undergoes issues of abrupt bank erosion and high sedimentation, stream displacement, soil fertility degradation occurs throughout the basin has raised the risk of soil erosion.

Kosi watershed situated in Almora and Nainital (central-eastern mountainous regions of Kumaon area of Uttarakhand) was preferred for this study.

OBJECTIVES

  • To delineate Kosi watershed using GIS and Remote sensing techniques
  • To determine morphometric characteristics of Kosi watershed
  • To prioritize Kosi watershed for watershed planning and management based on morphometric analysis using TOPSIS and VIKOR

Materials and Methods

Study area description:

Geographic location – 29°21’N to 29°52’N Latitude and 79°3’E to 79°52’E Longitude.

Total area - 1842.56 sq.km.

Altitude -363 m to 2732 m a.m.s.l.

Topography - Mountainous regions

Step-by-step

  • Primary data were obtained by SRTM (Shuttle Radar Topographic Mission) DEM imagery with a resolution of 30 m.
  • Morphometry: The geometry of the drainage basin and its stream channel system requires the measurements of three aspects:
  • The linear aspect of the drainage network,
  • The areal aspect of the drainage basin, and
  • The relief aspect of the channel network and contributing ground slopes.
  • Delineation: For delineation of the watershed, stream networks, morphometric study through QGIS 2.61 (Brighton), an open-source software, was used for the geospatial analysis.
  • Prioritization: The ranking of the Kosi Sub-Watersheds were done using TOPSIS and VIKOR MCDM Techniques

Results and Discussion

In the VIKOR model, sub-watershed 17, 13, and 19 with the highest results (0.982, 0.967 and 0.649) are most susceptible to erodibility, and sub-watersheds 6, 20, and 11 divulging lowest score (0.011, 0.019 and 0.088) have shown the least erosion susceptible.

Sub-watersheds 17, 13, and 19 with the maximum value (0.811, 0.803 and 0.784) of rank 1 to 3 are most vulnerable to erodibility, according to the outcome of prioritization of sub-watersheds in terms of their vulnerability to erosion using the TOPSIS process. Contrarily sub-watersheds 6, 20, and 11 have demonstrated the least vulnerability to erosion, with ratings of 0.241, 0.249, and 0.387, respectively.

Conclusion and Recommendation

  • In this study, the viability of sub-watershed prioritization based on morphometry of the Kosi watershed using Geographical Information System (GIS), also, VIKOR and TOPSIS multi-criteria decision making methods is demonstrated.
  • By the results, it has been settled that the sub-watershed SW17 falls into the top priority category. Therefore, the sub-watershed should be given importance for conservation.
  • Throughout, given the erosion vulnerability of the Kosi watershed, it is proposed that adequate mitigation measures be taken to minimize soil erodibility and reservoir dregs deposition, strengthen steep slopes against landslides, and reduce potential flood risk. Therefore, further analysis on estimation of gully erosion and sedimentation is also recommended.

References

  • Horton RE (1945) Erosional development of streams and their drainage basins; hydrophysical approach to quantitative morphology. Geological Society of America Bulletin 56(3):275–275
  • Strahler AN (1964) Quantitative geomorphology of drainage basins and channel networks. Section 4-II. In: Chow VT (ed) Handbook of applied hydrology. McGraw-Hill, New York

VIKOR and TOPSIS based Prioritization of Kosi Watershed using Geomatics

Purabi Sarkar, Pankaj Kumar, A. K. Sharma and S Sharma spurabi08@gmail.com

12 of 52

Healthy vs. Eutrophic Ecosystem

Introduction

Global Eutrophication Watch

An iterative app for coastal water managers to detects symptoms of coastal eutrophication solely using satellite-derived chlorophyll-a (CHL) concentration. The app can detect areas with severe symptoms of eutrophication and those with improving water quality and is relevant in tracking the progress against Sustainable Development Goals.

What was the motivation?

  • The accelerated degradation of coastal ecosystems associated with increasing anthropogenic nutrient loading is threating the integrity of many coastal ecosystems.
  • Evaluation of the impacts and extents of coastal eutrophication relies on extensive and intensive in-situ data collection.
  • The global eutrophication watch (de Raús Maúre et al. 2021, Nat. Commun.) was developed as a cost-effective planetary tool for coastal eutrophication monitoring.

How the assessment is done in the app

  • Indentify trends in satellite-derived chlorophyll (CHL) data at pixel level.
  • Discriminate waters based on the level of CHL composite from the most recent 3-year.
  • Combine the levels and trends and generate the six-pattern clastering shown below.
  • Embend this process inside an Earth Engine app, the global eutrophication watch

Results

Conclusion

Lessons Learned

  • The global eutrophication watch can identify eutrophication potential with the sole use of satellite-derived CHL and provides the initial screening of eutrophication for prioritized actions. It has the potential to be used as a global index of eutrophication.
  • The most difficult work is generating a CHL product that is suitable for the coastal waters under scrutiny.
  • With our default dataset, areas of severe symptoms of eutrophication (e.g., the Bohai Sea, the Gulf of Mexico and the Baltic Sea) are well captured by the app.
  • The global eutrophication watch is a relevant tool for coastal water managers and for tracking the progress against Sustainable Development Goals.

References

Globally consistent assessment of coastal eutrophication

Elígio de Raús Maúre1, Genki Terauchi1, Joji Ishizaka2, Nicholas Clinton2 & Michael DeWitt3

1Department of Research and Study, Northwest Pacific Region Environmental Cooperation Center (NPEC), Toyama, Japan. 2Institute for Space-Earth Environmental Research, Nagoya University, Nagoya, Japan. 3Google LLC, 1600 Amphitheater Parkway, Mountain View, CA, USA. ✉email: eligiomaure@gmail.com

Trends in annual CHL max based on Sen’s slope method (Sen, 1968) at 90% significance level. Polar regions with a few observations (< 70% of the study period) were masked.

  • Blue: CHL < 5 mg m-3 (open ocean mostly)
  • Red: CHL ≥ 5 mg m-3 (coastal waters mostly)
  • Eutrophic potential waters: HD, HN and HI
  • Eutrophication potential waters: HI and LI
  • From a composite of the most recent 3-year of the analysis period we split the waters into high-low CHL levels

Baltic Sea

Gulf of Mexico

Comparative eutrophication assessment in the Bohai Sea: a (1998-2015) and b (1998-2019)

  • Dashed ellipses highlight the decrease in eutrophication potential waters (LI and HI). For details see (de Raús Maúre et al. 2021, Nat. Commun.).

Eutrophication assessment in the Baltic Sea (c) and Gulf of Mexico (f) based on the global eutrophication watch .

Spatial distributions of bottom hypoxia and anoxia over time (Carstensen et al. 2014)

  • Red: bottom oxygen <2 mg L−1, Black: bottom oxygen <0 mg L−1 for 1993 (d) and 2012 (e) (Carstensen et al. 2014).

A satellite image during summer 2006 showing high concentrations of phytoplankton in reds and greens (g) (Miller & Spoolman, 2014)

e

d

c

f

g

h

The world’s third largest oxygen depleted zone (after the Baltic Sea and the northwestern Black Sea). Oxygen < 2 mg L−1 (h) (Miller & Spoolman, 2014)

Conceptual diagram comparing a healthy system to unhealthy system exhibiting symptoms of eutrophication (Bricker et al. 2007)

The global eutrophication watch app on the left and the animation of eutrophication assessment steps in the Northwest Pacific region on the right

13 of 52

Introduction

COVID-19 Breakthrough Hospitalizations & Deaths

Data on COVID-19 breakthrough infections, when fully inoculated people contract the virus, is relatively new. There is some evidence that these cases are becoming more common as the more transmissible COVID-19 Delta variant surges. The US Centers for Disease Control and Prevention estimated that fully vaccinated people accounted for less than 3% of coronavirus hospitalizations nationwide and less than 1% of virus deaths. Looking at how many hospitalizations and deaths have involved fully vaccinated people is a common, but crude measure of how well the vaccines are working.[1]

Rationale

  • COVID-19 breakthrough infection data is becoming more readily available and should be highlighted for all as time goes on
  • The general public could better benefit from visualizing this data on a map to highlight the effectiveness of vaccination against COVID-19 and poorer outcomes such as severe disease & death
  • We visualized this data for a select group of states in the Southeastern US to place the number of breakthrough cases in vaccinated persons in context of disease & death statistics for the wider COVID-19 pandemic

Methodology

The New York Times asked all 50 states and Washington, D.C. to provide data on breakthrough infections roughly spanning back to the first months of the COVID-19 vaccination campaign at the end of 2020 and beginning of 2021. Florida and Arkansas, while in the Southeastern US, did not provide data. [1]

Data from State health departments (breakthrough hospitalizations, breakthrough deaths and some total deaths), U.S. Department of Health and Human Services (hospitalizations), New York Times database of cases and deaths (some deaths) was used. [1] A Google sheet (shown below) was created highlighting the number of breakthrough COVID-19 hospitalizations & deaths from the NY Times data combined with publicly available US Census Bureau GIS state data. [3]

This combined information was then uploaded and styled accordingly for display using Google My Maps.

Results

Key: (State, # of Breakthrough COVID Deaths)

States are shaded based on the percentage of Breakthrough COVID-19 Deaths as compared to all COVID-19 Deaths. Darker shades indicate greater COVID-19 mortality (death) rates.

These results show us that the percentage of COVID-19 breakthrough deaths in select states in the Southeastern US are relatively low in the population of people who have been vaccinated against COVID-19.

View the full size version of the map here.

Conclusion

Recommendations

  • This data strongly supports existing research that receiving the coronavirus vaccine drastically reduces the chances of contracting and developing a more severe case of COVID-19 or a case that may result in death
  • This data should be reviewed again in the coming months to compare COVID-19 hospitalization and death rates among those who were not vaccinated against COVID-19 versus those were vaccinated in the US

References

Breakthrough COVID-19 Mortality

in Select Southeastern US States, 2021

Elizabeth Carter, elizabeth@publichealthmaps.org & Edward Paul Vallejo, edward@publichealthmaps.org

Limitations

  • Breakthrough case data was reported from 40 states and Washington, D.C., not all 50 states
  • The data collected covered the beginning of vaccinations being offered to the general public in late 2020 through mid-June to July 2021 depending on the State reporting data
    • While vaccines have done a remarkable job at protecting a vast majority of people from serious illness, more research and visualization is needed, especially as more breakthrough case data becomes available as the COVID-19 Delta variant surges and preventative measures are eased in certain parts of the US

14 of 52

  • SAFECITY crowdmaps personal experiences of sexual violence in public spaces.

  • This is aggregated as location based trends and visualised as hotspots on a map.

  • The aim is to make public spaces SAFE and equally accessible to all through
    • Individual awareness
      • Safe space to report
      • Information on rights
    • Community engagement
      • Data to take action
    • Institutional accountability
      • General awareness to make spaces safer
      • Data to make informed decisions

  • Vision - Safe spaces for all

Introduction

Sexual Violence is a pandemic

  • 1 in 3 women around the world experience sexual and gender based violence on an average
  • In India, every 15 mins, there is a rape reported somewhere in the country
  • 80% of these incidents are never documented due to socio-cultural norms and low trust in justice institutions and police
  • Women and girls are afraid to break their silence
  • This makes the issue “invisible” due to lack of data

  • Normalises the sexual violence because it happens so regularly
  • Creates a culture of insecurity where women and girls have to prove their innocence & affects their mental health
  • Difficulty on the part of service providers and law makers in making informed decisions

Method

Results

Conclusion

References

  • https://webapp.safecity.in/

Safecity - a crowdmap for sexual and gender based violence

ElsaMarie DSilva, elsamarieds@gmail.com

Non verbal

  • Staring / Ogling
  • Taking pictures
  • Indecent gestures

Verbal

  • Commenting
  • Catcalls
  • Sexual invites

Physical

  • Touching / Groping
  • Stalking
  • Sexual Assault
  • Rape

The data can be used by:

Police

Added insights into specific issues and locations

School and college administration

New data set which they may not have

Prevention of sexual violence on campus

Elected representatives

Focussed spending on budgets

Legal advocacy

Influencing policy, amending legislation or advocacy for new legislation where required

INDIVIDUALS

- can make better choices regarding their own safety

COMMUNITIES

  • can rally themselves around the issue
  • demand for better accountability from institutions

INSTITUTIONS

  • new dataset to work with
  • better relationship with communities

35000

reports

10 country chapters

1 Million

Citizens

Engaged

5 Police

Partners

40000

People trained

2000 Youth leaders

Achievements

15 of 52

Introduction

What is Accessible Life?

Accessible Life in Google Earth

is a visual search system that allows you to explore the earth to discover accessible outdoor locations.

The search - completely visual - takes place crossing successive layers until you reach the area you are looking for.

Accessible Life is a collective Map, built by contributors like us who want to help others.

Our Goals

  • Pivot to an accessible and scalable platform. Create a platform that can adapt to the number of contributions and remain fast and easy to navigate at the same time, thanks to the ability to add new layers based on the number of contributions
  • Focus on giving to every user the opportunity to evaluate. Not to decide for the user. Give them the possibility to decide if a place is Accessible according to their needs

How I Did it

I started by creating a standard structure in Google Earth, replicating it to create several interconnected layers.

In the next step I prepared the standard structure for Google My Maps, so I worked to have a coherent set of instructions to give to contributors to ensure a standard and consistent result. When everything was ready, I contacted some friends Local Guides, to see if they were able to create a map, applying the required standard, without further instructions.

Our tools

  • Google Earth: the search engine to which all maps created by contributors are linked
  • Google My Maps: The tool with which the Map of each single Accessible Outdoor Place is built
  • Google Maps: to provide links to all services located in the place
  • Google Street View: to give to the user the possibility of an immersive visit

Results

Conclusion

Where we are, Where we want to go

  • Through Accessible Life we are crowdsourcing data from volunteers who are doing this work manually, collecting information and putting them together in a rational way
  • I am quite sure that in the future, through Earth Engine and other Google Programs (I am thinking about Maps, but also Photos, Translator, Lens) Google will be able to
    • Scan the World
    • Identify the information
    • Organize them in a functional way

References

Ermes Tuon Poster: Accessible Life in Google Earth

Ermes Tuon, ertuon@gmail.com

We have currently mapped Accessible Places across four continents (Asia, Europe, North America, South America), and the map is constantly growing with new contributions.

The interaction with the contributors also allows to evolve and refine the system, adding new details to increase the ease of use by users.

16 of 52

Introduction

Introduction

What inspires me: The successful project from Ermes Tuon (ertuon@gmail.com) working on Google Earth inspires me to learn, to open to a new perspective in different ways, to challenge, and to participate.

My Goals

We have a beautiful State, Michigan USA, with so much to offer. If I can add more info about accessibility via Google Earth to make it just a click away I will. Let’s have the individual with the disabilities to determine if the location or parks is accessible for him/herself.

How I Did it

How I did it?

  • Decide what park your interested in.
  • Take photos that will be presented in the project
  • Get as many brochures about the park, the info can also be acquired from the internet.
  • Create and add the info to the Map, do the outline, name it, add photo into the point of interest
  • Add My Map into the Google Earth

The tools to use:

- Google My Maps: is the tool to use to create

the map

- Google Earth: is the search engine where my

maps are linked to.

Results

Conclusion

Conclusions:

This will be a wonderful way to access location with Google Earth and I strongly believe that in 5 years or more to the future, Google Earth will be the hot spot in the technology industry. A lot of works but fun to do!

Erna LaBeau Poster: Accessible Life in Google Earth

Erna LaBeau: ernalabeau@gmail.com

Everyone who has the link will be able to see the beauty of the parks in Google Earth. Every Point of Interest will be provided with photos and explanations. This way the viewer can imagine what to expect if they are to visit the location.

17 of 52

Introduction

Climate change awareness

We wanted to educate and support our communities on the growing challenge of climate change and urban resilience. This include increased vulnerabilities to flooding and rising temperature. In 2019, we developed a Nature-based Climate Adaptation Programme for the urban areas of Penang Island (first in Malaysia). We supported the programme through data-driven approach in our climate research and demonstrative projects which include our study of Urban Heat Island effects in Malaysian cities.

Our goals

  • Support our nature-based climate adaptation programme through empirical evidences
  • Educate and building community awareness on climate change
  • Communicate impacts in a manner that communities can easily understand

How We Did It

We chose five highly developed and populated cities in Malaysia for the study to assess the change in their surface temperatures between two different periods. Using GIS and Google data platforms, we were able acquire open-source satellite data captured by the United States Geological Survey (USGS).

Impacts and Implications

Conclusion

Lessons learned

  • Open-source data and science-driven evidences can move mountains
  • Effective visualization can close the gaps of understanding on how dire climate change impacts are
  • Urban researchers and strategists are given a new lens in understanding how different urban conditions contribute to UHI effects
  • Data from Google Engine can assist the quantification of ecological and economic benefits and values of urban forests and climate-resilient street trees

References

  • Think City, 2019. Penang Island Nature-based Climate Adaptation Programme (Link)
  • Think City’s Urban Analytics Portal (Link)

Cities are becoming hotter

Fatin Abdilah, fatin.abdilah@thinkcity.com.my

Rose Mansor, rose.afrina@thinkcity.com.my

Think City study: Malaysian cities got hotter over the decades; KL had lowest increase, Ipoh the most

FRIDAY 5 MARCH 2021

Perak moves to reduce urban heat

MONDAY 8 MARCH 2021

Adaptation Fund will help Penang reach its green development goals

SUNDAY 14 MARCH 2021

Malaysia’s getting hotter. Can its leaders rise to the climate challenge?

THURSDAY 18 MARCH 2021

  • Our study gained local and international attention, and is widely publicized
  • In January 2020, the Penang climate programme won the Climathon Global Cities Award, an award given by EIT-Climate KIC (Link)

Step-by-step

  • Extract Landsat data from USGS via Google Earth Engine
  • Process the data in ESRI software
  • Migrate the data into Think City Urban Analytics portal
  • Create story panel to interpret the findings and its implication to the city and community
  • Incorporate Street View imagery to assist observation
  • Document how the data has helped provide the evidences related to climate change
  • Launch the study to the public

  • In October, 2021, the programme was endorsed at concept note stage by the Adaptation Fund board to receive US$10 million in funding for implementation (Link)

18 of 52

Introduction

ABMI’s Mission

The Alberta Biodiversity Monitoring Institute’s mission is to track changes in Alberta’s wildlife and their habitats and provide ongoing, relevant, scientifically credible information for decision makers, land stewards, and all Albertans. Part of this mission is to produce open source and openly available geospatial landcover products. The Alberta wetland inventory is the culmination of three major regional wetland mapping initiatives and was published in March 2021.

Objectives of Alberta wetland mapping

  • Use best available open source Earth Observation data, cloud computing and machine learning techniques
  • Map to the five major classes of the Canadian Wetland Classification System
  • Provide alternative to Alberta Merged Wetland Inventory, which is an amalgamation of former localized wetland mapping initiatives from different scales and data sources

Method

The wetland classification was based on open source Sentinel-1 and Sentinel-2 imagery compiled in Google Earth Engine, ALOS DEM 30 m topographic data, and ABMI’s “photoplots” : eight hundred 3 by 7 km photointerpretation blocks spaced across Alberta which were used to train the classification models. The processing was done in Google Earth Engine, SAGA GIS, ArcGIS, R, and Python. Different object oriented modelling approaches were used for each of the three major regions.

Classification Models

  • Rocky Mountains: SNIC segmentation at 10 m scale and Random Forest model in GEE
  • Southern Alberta: SNIC segmentation at 5 m scale and seasonal band math to identify temporary and permanent flooded potholes
  • Boreal Region: Convolutional Neural Network implemented in Python with the Keras deep learning library, to incorporate contextual and texturual factors into the modelling of extensive peatland complexes, swamps, and marshes

Results

  • 22% of Alberta was classified as water or wetland. The most common wetland type is the 8 million hectares of fen, mainly in the boreal region. There are 1.8 million hectares of swamp and 1.1 million hectares each of bog and marsh.
  • Class accuracy was evaluated with a reserved subset of the photoplot data and by comparison to other datasets. Wetland vs. upland is over 90% accurate, and the overall accuracy of the five wetland classes plus upland is ~86%. Individual class accuracies vary and are most accurate in southern Alberta where there are fewer wetland types.
  • The data, detailed methods, and code can be downloaded from www.abmi.ca - Data & Analytics - Data Download. The wetland inventory can be interactively viewed as a layer in the ABMI Mapping Portal.

Conclusion

  • Google Earth Engine ‘s cloud computing capabilities were key to the efficient compositing of 661,848 km2 of multi-year and multi-seasonal imagery, which in turn was key to achieving best possible wetland class accuracies
  • Varying scales and parameters of SNIC clusters helped us find the best delineation of wetland features in different natural regions
  • Individual wetland classes are challenging to distinguish from each other, but the CNN deep learning model did offer a significant advantage for boreal fen and marsh classes over a XGBoost shallow learning model. There is a GEE app to compare these results at https://abmigc.users.earthengine.app/view/cnn-xgb
  • The ABMI Geospatial Centre continues to work with GEE on projects to assess hydro-temporal variability of prairie potholes, algal blooms on shallow lakes, regeneration of cutblocks and mine sites, and an upland Alberta landcover classification.

References

  • DeLancey, ER, Simms, JF, Mahdianpari, M, Brisco, B, Mahoney, C., & Kariyeva, J. (2020). Comparing deep learning and shallow learning for large-scale wetland classification in Alberta, Canada. Remote sensing, 12(1), 2
  • DeLancey ER., Kariyeva J, Bried JT, Hird JN. (2019). Large-scale probabilistic identification of boreal peatlands using Google Earth Engine, open-access satellite data, and machine learning. PloS one. 2019 Jun 17;14(6):e0218165

The Making of the Alberta Wetland Inventory

Fiona Gregory, Evan DeLancey, John Simms, Agatha Czekajalo, Jahan Kariyeva fgregory@ualberta.ca

Please replace these image placeholders with images related your work or project, or remove completely if you want to use this space for text.

ABMI’s Photoplot Interpretation Data

Image Processing in Google Earth Engine

SNIC segmentation for isolated wetlands in the Rocky Mountains

SNIC segmentation in southern Alberta prairie pothole region

Convolution Neural Network (CNN) model for the Boreal Forest region

Collaborators

Map by Evan DeLancey, ABMI

Alberta, Canada

19 of 52

Rationale

Growing A World Wonder

“Sustainably managed and restored trees and forests in Africa’s drylands have the potential to reverse land degradation and therefore contribute to poverty reduction, food security, biodiversity and climate change mitigation and adaptation.” (Sacande et al., 2020) The Great Green Wall of Africa is a transcontinental afforestation project spanning 8,000km long and 15km wide. It was launched in 2007, lead by the African Union in collaboration with other partnering institutions such as the United Nations Convention to Combat Desertification. The project aims to plant trees across the Sahelian region (the route passes through 11 countries) to reverse desertification processes and combat Climate Change.

Methods

Due to the transcontinental scale of the project, Earth observation satellite data was necessary for this study. Remote sensing methods provided the tools to assess vegetation dynamics across such a large study site, and at the time no monitoring and evaluation research on this project using remote sensing data had been published.

Results

Key Findings

  • MSAVI demonstrates a positive linear regression trend for the Great Green Wall region from 2007-2020, but with only a small gradient (2.68 x 10-4), suggesting little visible greening has occurred.
  • Ethiopia and Eritrea demonstrated the highest significant greening trend of all the countries.

Conclusion

  • This study provides an open-source monitoring and evaluation tool for the Great Green Wall of Africa, accessible to anyone from institutional project leaders to smallholder farm owners, that can be run at any scale for any region, and that is repeatable at any point throughout the duration of the project.
  • Visible regreening from space is not yet evident, but this is to be expected given that the coarse resolution of MODIS Terra is unlikely to pick up growth of small tree saplings at a macro scale. The scale by which success is measured should be rethought, and more local scale microclimatic changes should be quantified and recognised.

References

A Novel Open-Source Remote Sensing Tool To Improve Monitoring and Evaluation of the Great Green Wall of Africa

Gemma Newbold, gemma@justdiggit.org

Challenges limiting progress towards the project’s ambitious targets

Based on a recent assessment, many challenges (financial, technological, biophysical, governance, and monitoring) to the implementation of the Great Green Wall of Africa have meant the project is not on track to meet its ambitious 2030 targets, which include:

  • Restore 100 million hectares of degraded land by 2030
  • Sequester 250 million tonnes of carbon by 2030
  • Meet 15/17 of the Sustainable Development Goals

Therefore, there was a research gap for a monitoring tool to provide temporal and spatial insights into the success of the project so far, and where more funding or technical expertise needs to be directed.

Benefits of the project

  • Carbon sequestration and climate regulation
  • Creating shade, cooling land surface temperatures
  • Increased evapotranspiration, driving the local hydrological cycle
  • Ecosystem restoration and increased biodiversity
  • Improved food and water security

Dataset

Vegetation Index

MSAVI

The Modified Soil Adjustment Vegetation Index, which was chosen for the analysis, performs better than NDVI in sparser canopies, as it corrects for soil background reflectance.

MODIS Terra

  • Data product: MOD09A1.006 Surface Reflectance
  • Data availability: 05-03-2000 until present
  • Provider: NASA
  • Temporal resolution: 8 day
  • Spatial resolution: 500m

Methodology

  1. Clip to region geometry
  2. Extract time series data
  3. Export to R for analysis

A ‘greened’ pixel is defined, for the purpose of this analysis, as a pixel which has a positive difference in MSAVI value when subtracting its 2007-2009 average from its 2018-2020 average. This greening could have occurred from natural causes or human intervention.

  • Water (precipitation and soil moisture) is critical. Crossing the critical threshold past which hydrological cycling feedback loops set in, which are able to self-sustain vegetation cover across the region and accelerate its growth, will be essential to the project’s long-term success.
  • GEDI, a recently launched mission to provide global LiDAR coverage, offers future potential for accurate monitoring of above ground vegetation carbon content of the Great Green Wall of Africa, which should be utilised for future research.

20 of 52

Background

The demographics of trees

Many studies have explored the relationship between forest and tree cover (Riley and Gardiner 2020) and green spaces (Dai 2011) to different demographics such as race (Lu et al., 2021) and income (Holt and Borsuk 2020). In this study, I looked at the relationship between forest cover, race and income in my home state of Virginia.

Typically, studies conclude that people who are white, asian, and/or higher income (Pearsall and Eller 2020) reside in areas with higher percentages of trees and green spaces. As of 2020, Virginia has a population of over 8.6 million people; of which 67% are White, 19% are Black, and 6% are Asian [1]. In 2019, the overall median household income for the state was $74,222 [2].

Objectives

I wanted to look at how forest cover relates to race and income in Virginia. Do the relationships that past papers have shown hold up here? If not, how much do these relationships differ?

My objective is:

  • What does the relationship between race and forest cover, and income and forest cover look like for census tracts in Virginia?

How to

To determine the relationship between tree cover, race, and income, the following datasets were used:

  • The Hansen Global Forest Cover raster dataset for 2019 downloaded from Google Earth Engine [3]
  • US Census Tract boundaries
  • US Census race and median household income information for 2020

The Forest Cover raster data was summarized by census tracts to create a total forest cover % by census tract. This way I could see the total % forest cover, race, and income for each census block.

The relationships between these three variables were done by running correlations between the three variables using the ggcorrplot [4] and ggpubr [5] libraries in R.

The correlation of these variables tells me the type of relationship (positive or negative) that race and income has against forest cover for each census tract.

Results

Conclusion and next steps

Final thoughts and where to go from here:

  • This study showed that the relationships found in past studies are found in the state of Virginia at the census tract level.
  • My results showed that mean forest cover (%) had the highest correlation with White identifying populations. Mean forest cover (%) had negative correlations with median household income, and Black and Asian identifying populations.
  • If I were to continue this study, I would look into the relationship of mean forest cover to multiple other variables at once.
  • I would also like to map correlations between forest cover, race, and income to show what census tracts have positive or negative correlations.
  • I would also like to be able to compare these relationships with other studies done in Virginia, and other states!

There are only a few of the ways I would like to improve upon this study and am very interested in hearing your perspective!

References

[1] Virginia population and race percentages: https://worldpopulationreview.com/states/virginia-population

[2] Virginia overall median household income: https://www.census.gov/quickfacts/fact/table/VA/AFN120212

[3] Hansen Global Forest Cover rasters: https://earthenginepartners.appspot.com/science-2013-global-forest/download_v1.7.html

[4] Ggcorrplot: Alboukadel Kassambara (2019). ggcorrplot: Visualization of a Correlation Matrix using 'ggplot2'. R package

version 0.1.3. https://CRAN.R-project.org/package=ggcorrplot

[5] Ggpubr: Alboukadel Kassambara (2020). ggpubr: 'ggplot2' Based Publication Ready Plots. R package version 0.4.0.

https://CRAN.R-project.org/package=ggpubr

The Demographics of Trees in Virginia

Gloria Desanker, M.S. | gdesanker@gmail.com

Principal Consultant, Map Nerd Consulting

This flowchart shows the general process I used to create my final map of race vs forest cover and income vs forest cover.

Mean forest cover had the highest correlation with White identifying populations.

% Mean Forest Cover by Census Tract

Population by Census Tract

Median Household Income by Census Tract

21 of 52

Introduction

Why need to create a landslide imagery database

Landslides are common natural disasters around the globe. Understanding the accurate spatial distribution of landslides is essential for landslide analysis, prediction, and hazard mitigation. So far, many techniques have been used for landslide mapping to establish landslide inventories. However, these techniques either have a low automation level (e.g., visual interpretation-based methods) or a low generalization ability (e.g., pixel-based or object-based approaches); and improvements are required for landslide mapping.

What was our goal

  • Create a landslide dataset with annotationed labeling that can be used to develop landslide deep learning model
  • We need to collect landslide imary and create an easy annotation tool for interactive landslide labeling

Method

We used Google earth engine as backend to enable users to search on NAIP imagery on Google map and discover/predict the landslide by drawing the bounding box in the map interface.

Our tool pipeline

  • Our deep learning model is developed on the cloud engine
  • Once user submit the area of interest from the google map, it will upload region though the google Earth Engine and upload NAIP imagery to Cloud Storage.
  • Deep learning model can directly access the Cloud Storage uploaded tif and predict the landslide detect areas

Results

Our web portal is developed based on the google cloud platform. It provides a user-friendly interface to improve landslide mapping efficentcy. The deep learning model guides users’ attention to landslide- suspicious area in the image. Users can review and correct machine labeled landslides using the labeling tool.

Conclusion

  • We manually identified 2,509 landslide points and collected 496 “bi-temporal” satellite image pairs from Google Earth and Earth Engine.
  • Each image pair consists of a pre-event image before the landslide event and a post-event image after the event.
  • We are partnered with the United States Geological Survey (USGS) and Federal Highway Administration (FHWA) now to use our tool/dataset/model identifying, collecting and predicting landslide events in the North America region.

Utilizing an interactive AI-empowered web portal for landslide labeling

Guanlin He (gbh5146@psu.edu) ,Te Pei, Savinay Nagendra, Srikanth Banagere Manjunatha, Daniel Kifer, Tong Qiu, and Chaopeng Shen

Department of Civil and Environmental Engineering, Pennsylvania State University

Department of Computer Science, Pennsylvania State University

Please replace these image placeholders with images related your work or project, or remove completely if you want to use this space for text.

Please replace these image placeholders with images related your work or project, or remove completely if you want to use this space for text.

.

22 of 52

Introduction

Mapping Habitat Loss

Agricultural expansion is a major driver of habitat loss and threat to biodiversity in sub-Saharan Africa (Tilman et al., 2017). Accurate land use maps help conservation planners monitor ecosystem health and manage protected areas. In Angola, however, thousands of landmines have made mapping agricultural expansion on foot impossible. And a dynamic savanna ecosystem, structured by annual fires, means that well-vetted techniques used to map habitat loss in forests do not transfer. Google Earth Engine’s global imagery database and machine learning algorithms provided a unique opportunity to create a new, more accurate habitat loss map for the Angolan miombo woodlands ecosystem.

What was our goal?

  • Create a new map of habitat loss in Angola’s miombo woodlands, a savanna ecosystem 5x the size of England

  • Above Figure: The Angolan miombo woodlands
  • Left Figure: Extent of the miombo woodlands (gray) in Angola, and areas experiencing fire (red) in 2018

Method

Methods

  • 5000 training points manually selected across 11 sub-regions of Angola’s miombo woodlands

  • Expert knowledge used to separate points into 4 classes (see Figure to right)

  • Landsat 8 imagery used within GEE to classify pixels across the landscape

  • Results verified through reference to very high resolution imagery in Google Earth

  • Post-war expansion of agriculture quantified within GEE and interpreted using knowledge of Angolan ecology and politics

Results

Figure: Distribution of woodland (dark green), grassland (light green), agricultural areas (pink) and urban areas (red) in the Angolan miombo woodlands, mapped using GEE with 86% accuracy.

Agriculture has expanded significantly (by 5%) since 2013, driving loss of savanna woodland and grassland habitat. Rapid expansion is occurring (1) along major roads (2) around regional cities and (3) along an agricultural frontier advancing northeastward from the Angolan highlands—the large agricultural belt in the region’s west.

Conclusion

Key Points and Opportunities for Future Study

  • Habitat loss in Angola’s savanna woodlands is occurring in predictable patterns, along roads, frontiers, and near cities
  • GEE can be used to create new, more accurate maps of agricultural expansion in dynamic, difficult to reach ecosystems
  • Better maps make possible future studies linking habitat loss to political dynamics such as war, post-conflict resettlement, and infrastructure reconstruction
  • Better maps allow us to relate habitat loss to environmental changes such as fire, biodiversity loss, and woody encroachment

References

  • Tilman, D., Clark, M., Williams, D.R., Kimmel, K., Polasky, S., Packer, C., 2017. Future threats to biodiversity and pathways to their prevention. Nature 546, 73–81
  • Huntley, B.J., Ferrand, N., 2019. Angolan Biodiversity: Towards a Modern Synthesis, in: Huntley, B.J., Russo, V., Lages, F., Ferrand, N. (Eds.), Biodiversity of Angola: Science & Conservation: A Modern Synthesis. Springer International Publishing, Cham, pp. 3–14.

Mapping Post-Conflict Habitat Loss in Angola’s Woodlands

Imma Oliveras Ph.D., Ty Loft ty.loft@bnc.ox.ac.uk

Please replace these image placeholders with images related your work or project, or remove completely if you want to use this space for text.

A chameleon crossing the road in the miombo woodlands. Photo credit: Ty Loft

23 of 52

Introduction

A cash crop amongst many...

West Africa has lost 90% of its original moist forest and what remains is heavily fragmented and degraded (Leach and James, 2000). Cocoa farms, amongst others, such as oil palm, rubber and coconut are the major drivers of deforestation in Ghana and Côte d’Ivoire (Ruf, 2015). Ghana has lost more than 2.5 million hectares (Mha) (33.7%) of its forest since the early 1990s (Fuglie et al., 2013). The statistics for Côte d’Ivoire are even worse, as it has experienced rapid deforestation since the mid-1950s (Barraclough and Krishna, 2000) and this is partially due to the increase in cocoa demand (Nalley et al., 2014). It was recorded that between 2000 and 2013, cocoa plantations in West Africa were responsible for 57% of annual global expansion, and in 2013, 6.3 Mha was allocated to cocoa cultivation in the region (Ordway et al., 2017).

Protected areas (PAs) constitute a major conservation tool for protecting forests and the biodiversity they shelter. In tropical Africa, management effectiveness of several PAs is deficient, and that PAs continue to be exposed to threats such as wildlife hunting, logging and agriculture.

Cocoa farming is one of the agricultural activities threatening PAs in tropical Africa.

What was my goal?

  • Generate and present a cocoa thematic map for Côte d’Ivoire and Ghana for the year 2019 using a state of the art cloud-based computing platform and free and open access satellite data;
  • Estimate the area of cocoa plantations based on the developed thematic maps; and
  • Assess the extent of cocoa farming within and around protected areas.

How I Did It

An approximate 20 degree belt either side of the Equator with humid tropical climates with regular rain and a short dry season is favourable to the growth of cocoa (Kyei et al., 2011). As shown in this Figure, climatically the most suitable cocoa areas in Ghana (GH) are mainly in the Eastern, Central, Ashanti, Western and southern Brong-Ahafo regions, while in Côte d'Ivoire (CIV) they are mainly in Sud-Comoé (Comoé since 2011), Agnéby (Lagunes since 2011), Moyen Comoé (Comoé since 2011) and Sud-Bandama (Bas-Sassandra since 2011) districts (Läderach et al., 2013). Accordingly, we focused our mapping efforts on these regions and districts and neighbouring regions.

Result & Discussion

The area of cocoa recorded for cote d’Ivoire was 3.69 Mha and 2.15 Mha for Ghana. The cocoa plantations detected in close proximity to Protected Areas (PAs) may correspond to the cocoa agroforestry systems which have been suggested as a management strategy to reduce the influence of surrounding land-use activity on PA biodiversity. Although our study does not enable the direct investigation of how effective this strategy is in mitigating the effects of land-use activities on PA biodiversity, we clearly show that this strategy is ineffective in preventing the encroachment of cocoa plantations into PAs. The PAs housing cocoa plantations are concentrated in the southern parts of Côte d’Ivoire and Ghana, corresponding to the cocoa belt, and 105 PAs have cocoa plantations exceeding 35% of their area.

Conclusion

  • Cocoa plantations can be mapped using freely available radar (Sentinel-1) and optical (Sentinel-2) imagery and their derivatives (vegetation indices, texture measures);

  • Cocoa farms in Côte d’Ivoire and Ghana have a broad spatial distribution;

  • Cocoa farms have largely encroached into PAs.

Detecting cocoa plantations in Côte d’Ivoire and Ghana and their implications on protected areas

Itohan-Osa Abu, Zoltan Szantoi, Andreas Brink, Marine Robuchon, Michael Thiel

itohan-osa.abu@uni-wuerzburg.de

For more information please scan the barcode

The dataset consists of median composites of S-1 (VV and VH) and S-2 (13 bands) for the year 2019 (Figure 2). The size of the median window was set to the twelve-month period lasting from 1/1/2019 to 31/12/2019 because it was the minimum window size that could produce a cloud-free composite for S-2 in the study area. After initiating these steps, the S-1 composite was created for deriving S-1 derivatives and layer stacking with S-2 composite’s derivatives for the classification. Based on the S-2 composite, the NDVI calculation and Tasseled Cap Analysis (TCA) has been performed using functions written in GEE. Oil palm plantations (Descals et al., 2019), settlements (JRC Global Human Settlement Layer) and water bodies (JRC Global Surface Water Explorer) were masked out from the S-1 and S-2 composite and the derivatives.

Hansen, Matthew C., et al. (2013)

The climatically most suitable cocoa regions in Ghana and districts in Côte d’Ivoire (Läderach et al., 2013).

Cocoa plantations map for Côte d’Ivoire and Ghana representing the year 2019.

Fig. 8. Estimated area of cocoa plantations (%) within protected areas. NP = National Park, FR = Forest Reserve, RR = Resource Reserve, NR = Not Reported. 1. Marahoué NP, 2. Niégré FR, 3. Dassieko FR, 4. Monogaga FR, 5.Mabi FR/Yaya FR, 6. Séguéla FR, 7. Rapide Grah FR, 8. Kani-Bandaman NR, 9. Azagny NP, 10. Sassandra FR, 11. Haut De, NR, 12. Bolo FR, 13. Bossematie FR, 14. Mont Peko NP, 15. Moyenne Marahoué NR, 16. Koba FR, 17. Port Gautier FR, 18. Bouafle FR, 19. De FR, 20. Ngadan Ngadan FR, 21. Banco NP, 22. Ile Ehotilé NP, 23. Bia Tawya FR, 24. Tai NP, 25. Anhia NR, 26. Bia NP, 27. Tano Ofin FR, 28. Asenanyo FR, 29. Bia North FR, 30. Manzan FR, 31. Sukusuki FR, 32. Ankasa RR.

24 of 52

The Why

Alberta’s Dynamic Landscapes

The landscapes of Alberta, Canada, are ever-changing. In addition to natural disturbances (e.g., wildfire, insect defoliation), human land uses like resource extraction, agriculture, and urban expansion lead to continuous changes. Up-to-date mapping and monitoring are key to effective land use planning, and sustainable resource management.

Existing and comprehensive data on human footprint (HF) in the form of province-wide digital Human Footprint Inventories [1] are a valuable source of information (see map to right). They treat all features equally, however, regardless of the temporary nature of some features (e.g., forest harvest areas).

Our goal is to enhance current HF data for the province of Alberta with remote sensing-based information on vegetation recovery.

How We Did It

Our Results

Key Takeaways

References

  • Alberta Biodiversity Monitoring Institute and Alberta Human Footprint Monitoring Program Wall-to-Wall Human Footprint Inventory 2018 2020, 145p.
  • Key, C.H.; Benson, N.C. Landscape assessment: Sampling and analysis methods. USDA For. Serv. Gen. Tech. Rep. RMRS-GTR-164-CD 2006, pp. 1–55, doi:10.1002/app.1994.070541203.

Mapping post-disturbance recovery on human footprint features in Alberta, Canada

Jen Hird, jennifer.hird@ucalgary.ca, @JNHird

  • Number, year(s), and magnitude(s) of spectral disturbance events
  • Spectral recovery (as % of pre-disturbance signal) achieved after 5 years
  • Number of years required to reach 80% spectrally recovered
  • Total spectral recovery for most recent HF inventory

Earth Engine & the Landsat Archive

We leveraged the long-running Landsat archive using Google’s Earth Engine platform to extract spectral recovery trajectories for several types of polygonal HF features in Alberta. In short, we:

  • Combined Landsat 5, 7, and 8 images into annual, summer-time median pixel composites of the Normalized Burn Ratio (NBR) [2]
  • Detected spectral disturbances, extracted per-pixel spectral recovery metrics
  • Summarized these metrics for selected HF features. Our extracted metrics included:

Forest Harvest Areas (Alberta)

Well Sites (Oil Sands)

Reclaimed Mines (Oil Sands)

The majority of harvest areas in our provincial dataset (n = 57.8K; harvested 1989-2013) have recovered their pre-harvest NBR spectral signals in 2018.

Reclaimed mine areas in our Alberta Oil Sands dataset (n = 377) show highly variable levels of spectral recovery in 2018.

Active well sites in our Alberta Oil Sands dataset (n = 545) show some variety in levels of spectral recovery in 2018, but a peak around 65-86% spectrally recovered.

Abandoned well sites in our Alberta Oil Sands dataset (n = 3427) show less variability in spectral recovery in 2018, with an even higher peak between 65-86% spectrally recovered.

𝜇 = 72.8%

𝜎 = 22.8%

𝜇 = 105.2%

𝜎 = 20.9%

𝜇 = 76.3%

𝜎 = 25.3%

𝜇 = 66.7%

𝜎 = 24.7%

We Learned That:

Spectral recovery varies for different HF features, but shows some expected patterns (i.e., lower for active wellsites, higher for regenerating harvest areas).

Main Contributions:

  • An adaptable, scalable, and easily repeatable workflow for characterizing spectrally-based recovery on HF features with Landsat imagery
  • A public dataset of HF spectral recovery for Alberta harvest areas (see www.abmi.ca)

What’s Next? Analyses of environmental driving and limiting factors, and comparisons with ground observations to better understand our recovery metrics

25 of 52

Introduction

Coastal wetlands

The long-term fate of the mud dominated coastline of Suriname, part of the Guianas coastal system stretching from the Amazon river to the Orinoco delta, is influenced by migrating subtidal mudbanks. Their alongshore movement causes a cyclic instability of alternating erosion and progradation phases. For climate resilient management strategies, that need to to account for changes that might occur in the coming decades, it is critical to incorporate spatial and temporal variability of coastline dynamics that reflect the complex interplay between controlling factors such as mudbank migration.

Mudbank cyclicity

  • Mudbanks are associated with extreme spatiotemporal variations in suspended particulate matter concentrations that originate from the Amazon river and are transported along the coast by the Guiana current.
  • Mudbanks can reach up to 30 kilometers offshore and 20 kilometers alongshore.
  • Enhanced sediment deposition and increased wave damping, associated with mudbanks presence, provide a window of opportunity for mangrove species to rapidly colonize large intertidal surfaces.
  • When a mudbank after 10 – 15 years has migrated further, the coastline and mangrove ecosystems are again susceptible to enhanced erosion as a result of increased wave activity.
  • Boundaries between mudbanks and adjacent heterogeneous coastal features are inherently diffuse and change rapidly.

Method

Remote Sensing

We use a remote sensing approach (de Vries et al., 2021) to quantify the influence of mudbank migration on local coastline dynamics, along the entire coast of Suriname between 1985 and 2020 from medium resolution Landsat images that are available in Google Earth Engine.

For subtidal features, resuspension of mud and migration processes are responsible for the spatiotemporal variability of their footprints. Abrupt changes in mud abundance are thus considered as fuzzy transitions that correspond to mudbank boundaries. Accordingly, image analysis techniques that involve semi-automatic unmixing of cover fractions can be used for the analysis of these gradients. Simultaneously, the NDWI index is used to detect the interface between land and water with an Otsu thresholding approach (Donchyts et al., 2016).

Results

  • Migration of 6 -8 subtidal mudbanks causes on average expansion of 25 m/yr during mudbank presence, and 8 m/yr retreat of the coastline when mudbanks has migrated further alongshore.
  • High peak fractions in unmixed mud abundance maps indicate the presence of mudbanks with clear shifts from east to west that can be associated with their migration direction- and speed.
  • Coastline positions estimates, often coinciding with the mangrove fringe, are on average accurate within 50 meters.
  • Accretion and erosion patterns can be distinguished at relative local scales (10 - 30 km).

Conclusion

  • Standardized and objective method applied here results in comparable estimates of mud abundance that can be used to estimate the presence of subtidal mud banks.
  • Coastline change analysis for Suriname with unprecedented spatial coverage and temporal resolution.
  • Spatial- and temporal distribution of coastline changes significantly different between regions.
  • Yet, not all observed coastline changes can be attributed to the migration of mudbanks
  • Suggesting that when local settings are altered, erosion during interbank phases can be significantly larger than the accretion rates in the preceding mudbank phase.

References

  • de Vries, J., van Maanen, B., Ruessink, G., Verweij, P. A., & de Jong, S. M. (2021). Unmixing water and mud: Characterizing diffuse boundaries of subtidal mud banks from individual satellite observations. International Journal of Applied Earth Observation and Geoinformation, 95, 102252.
  • Donchyts, G., Baart, F., Winsemius, H., Gorelick, N., Kwadijk, J., & Van De Giesen, N. (2016). Earth's surface water change over the past 30 years. Nature Climate Change, 6(9), 810-813.

Coastline Dynamics Controlled by Migrating Subtidal Mudbanks from Remote Sensing Images in Google Earth Engine

Job de Vries

Unsupervised Decision Tree (UDT)

  • All individual (>1000) Landsat images between 1985 and 2020
  • Semi-automatic linear spectral unmixing of mud, water and vegetation fractions based on end-member signatures extracted from these images.
  • Intermediate outputs, including a binary land masks and filtered mud fraction maps.
  • Presence and absence estimates of mudbanks.
  • Characterize multitemporal coastline position changes for equally spaced, shore-normal orientated transects.

Faculty of Geosciences

Departhment of Physical Geography

j.devries4@uu.nl

https://www.uu.nl/staff/JdeVries5

Acknowledgements

The MangroMud project is funded by NWO-WOTRO (W 07.3030.106) and Utrecht University.

https://www.nwo.nl/en/projects/w-07303106

Figure 3.The presence and absence of mudbanks in relation to observed coastline changes in Suriname between 1986 and 2020.

Figure 1.The coastal area of Suriname with the different coastal sections separated by the two major rivers.

Figure 2. The applied workflow to determine the coastline position and mudbank presence.

Figure 4. Distribution of the rates of coastline change between 1985 and 2020 for all positions along the coast of Suriname that are either fronted by a mudbank or not.

26 of 52

Introduction

UNI Tallgrass Prairie Center Programs

The University of Northern Iowa (UNI) Tallgrass Prairie Center (TPC) is a leading native prairie proponent in the midwestern United States and has a demonstrated 30-year commitment to prairie reconstruction, restoration, management, and advocacy. The primary programs of the Center are Research and Restoration, Iowa Roadside Management (IRM), Plant Materials, and Prairie on Farms (POF).

Overview

  • Collaboration between UNI GeoTREE Center (Geography) and the TPC
  • Developed dynamic and interactive virtual resources to tell the story of the TPC POF and IM programs using a suite of Google technologies

Method

In collaboration with the TPC, the UNI GeoTREE Center planned collection of 360° photos at a variety of Prairie on Farms and Iowa Roadside Management locations. GeoTREE Center student research assistants collaborated with TPC Americorps members to leverage those photos and Google Earth technologies to build a suite of virtual resources.

Methods

  • Published 191 photos to Google Street View
  • Using Google Earth Studio developed virtual flyover videos for POF and IRM programs
  • Leveraging 360° photos and Google Earth built virtual tours of POF and IRM programs

Results

There were 191 photos published to Google Street View while multiple virtual resources were developed for both the Iowa Roadside Management and Prairie on Farms programs. These virtual resources include POF and IRM flyovers and IRM Google Earth Projects. The GeoTREE Center also developed a custom web mapping application using Google Maps JavaScript API and open source 360 photo viewer.

Conclusion

Accomplishments and Acknowledgements

  • This project led to the development of greatly improved virtual resources that are being used by the TPC to tell their story in a variety of ways
  • This project was funded by the Iowa Department of Agriculture and Land Stewardship
  • Contributors at UNI TPC: Kristine Nemec, Andrew Dunham, Laura Fischer Walter, Ethan Evans, Andy Olson, Rowan McMullent Chang, Paige Shafer
  • Contributors at UNI GeoTREE Center: Nathan Huffman, Zach Fuller, Ryan Lange, Greg Klocke, Paula Carvalho de Castro, Pratik Poudel, Grant Burke

Google Earth Technologies to Communicate Iowa Prairie Stories

John DeGroote, john.degroote@uni.edu

27 of 52

Introduction

Grasslands are important for global biodiversity, food security, and climate change analyses. Yet, grassland vegetation monitoring at spatial and temporal resolution relevant to land management (e.g., ca. 30-m, and at least annually over long time periods) is challenging due to, among others, often limited data availability.

Our new approach bridges this gap, allowing to map short- and long-term changes in grasslands at 30-m resolution based on the entire Landsat record available in Google Earth Engine,

We tested the approach in the grasslands in the Caucasus ecoregion.

Method

Results

Presented methodology and results are part of Lewińska et al. (2020) and Lewińska et al., [in revision].

Conclusion

References

  • Kennedy, R.E., Yang, Z., Gorelick, N., Braaten, J., Cavalcante, L., Cohen, W.B., Healey, S., 2018. Implementation of the LandTrendr algorithm on Google Earth Engine. Remote Sens. 10, 1–10. https://doi.org/10.3390/rs10050691
  • Kong, D., Zhang, Y., Gu, X., Wang, D., 2019. A robust method for reconstructing global MODIS EVI time series on the Google Earth Engine. ISPRS J. Photogramm. Remote Sens. 155, 13–24. https://doi.org/10.1016/j.isprsjprs.2019.06.014
  • Lewińska, K.E., Hostert, P., Buchner, J., Bleyhl, B., Radeloff, V.C., 2020. Short-term vegetation loss versus decadal degradation of grasslands in the Caucasus based on Cumulative Endmember Fractions. Remote Sens. Environ. 248. https://doi.org/10.1016/j.rse.2020.111969
  • Lewińska, K.E., Buchner, J., Bleyhl, B., Hostert, P., Yin, H., Kuemmerle, T., Radeloff, V.C., Changes in the grasslands of the Caucasus based on Cumulative Endmember Fractions from the full 1987-2019 Landsat record, in press in SRS, https://doi.org/10.1016/j.srs.2021.100035

Changes in the grasslands of the Caucasus based on

Cumulative Endmember Fractions from the full Landsat record

Katarzyna Ewa Lewińska*, Johanna Buchner, Benjamin Bleyhl, Patrick Hostert, He Yin, Tobias Kuemmerle, and Volker C. Radeloff

* corresponding author lewinska@wisc.edu | @kelewinska | http://silvis.forest.wisc.edu

Results

The approach is based on the Cumulative Endmember Fractions (CEF), i.e., annual sums of monthly ground cover fractions of: soil, green vegetation, non-photosynthetic vegetation and shade (Lewińska et al. 2020).

Processing consists of the following steps:

  • Spectral Mixture Analysis of each Landsat scene available in GEE for a selected study area;
  • derivation of monthly composites for each fraction;
  • inferring of missing monthly data using the Whittaker filter, as implemented in Kong et al. (2019);
  • aggregation the monthly time series of endmember fractions to annual sums (CEF);
  • adding shade and non-photosynthetic vegetation for simplicity;
  • LandTrendr (Kennedy, 2018) analysis to identify temporal changes in green vegetation CEF;
  • quantification of shift among ground cover fractions, hereafter change pathways.

Changes are classified into negative change pathways:

  • desiccation
  • green vegetation loss
  • dry vegetation loss

and positive change pathways:

  • greening
  • revegetating of green fraction
  • revegetating or dry fraction.

Grassland vegetation in the Caucasus was highly dynamic during 1987-2019.

  • positive change pathways affected 32.7% of grasslands;
  • negative change pathways occurred in 20.9% of grasslands;
  • 8.4% of grasslands went through a sequence of positive and negative changes (in any order).

Short-term negative change pathways were most abundant in 1990s and 2010s, while positive short-term change pathways were the most common before 1997.

Changes in grasslands were not controlled by weather conditions and the effect of livestock varied among regions.

Our approach:

  • allows to map changes in grasslands based on ground cover fractions at 30-m resolution since the mid-1980s;
  • identifies short- and long-term changes providing unique spatio-temporal information on vegetation dynamics;
  • is physically based, ensuring a strong quantitative foundation to study grassland change processes;
  • is globally applicable yet management-relevant at local scales.

28 of 52

Why you should care about Hong Kong’s forests

Hong Kong’s Natural Beauty

  • 2/3 of Hong Kong's total land area exists outside of urban areas
  • Wildfires in Hong Kong burn around 5% of the natural landscape every year, which poses major risk to biodiversity, releases carbon into the atmosphere, and can cause landslides
  • As projected future climate change may cause longer and warmer dry seasons, the risk for future wildfires may be greatly increased
  • In order to understand how wildfires spread once ignited, clear information about vegetation type must be known

Goals:

  • Combine freely available 2016 LiDAR data and Sentinel-2 MSI data
  • Build a Random Forest classifier to differentiate between vegetation types
  • Classify different vegetation types for all of Hong Kong’s natural areas

Combining Freely Available Datasets

Pixel- versus Object-based Classification

Looking Forward

  • Vegetation mapping is only the first step to gaining a comprehensive understanding of how wildfires will continue to impact Hong Kong’s forests
  • Next, forest metrics such as biomass, stem density, and canopy bulk density will be estimated using a combination of LiDAR and optical data
  • Biomass equations need data specific to forest-stand type which will be derived through this classification methodology

Harnessing the Power of Google Earth Engine to Map Vegetation in Hong Kong

Katherine Strattman katstratt1@gmail.com, Jed Kaplan jkaplan@hku.hk

revise

First Classification attempt:

Second Classification attempt:

  • Pixel-based RF classification showed speckled pixels, especially in vegetated areas
  • Different vegetation types are mixed, and the shrubland and grassland is mis-classified as forest
  • Object-based RF classification showed improvements, but is likely overfitted
  • Grassland, shrubland, and forests are clearly delineated

Pixel-based Classification

Object-based

Classification

Sentinel-2 MSI

Image Stack

Pixel-based Random Forest Model

LiDAR-derived

CHM and DEM

Sampled Training Points

Categorical Land Use/Land Cover Points

Sampled Testing Points

Validation and Prediction

Mask all Non-vegetated Areas

Sampled Training Points

Sampled Testing Points

Object-based Random Forest Model

Low Quality Classification

Sentinel-2 MSI

Image Stack

Validation and Prediction

High Quality Classification

LiDAR-derived

CHM and DEM

98% Accuracy

84% Accuracy

29 of 52

Introduction

Drought in the Navajo Nation

The Navajo Nation is the largest sovereign tribal nation in

the U.S. in total land area, with boundaries in Arizona,

New Mexico, and Utah.

On the Navajo Nation, they have observed increased drought, exacerbated by warming temperatures and precipitation deficits, which is adversely affecting their lives, including their farmland (>212,000 acres) that produce $92 million in agricultural products annually.

The land size and the region’s diversity, ranging from deserts to snowy forests, makes it challenging to identify who was most impacted.

Objective

Generate spatially resolved drought information that supports distribution of emergency funds to regions that are more severely affected by drought

Method

This project is a collaboration between the Navajo Nation, the Desert Research Institute (DRI) and NASA’s Western Water Applications Office. In order to meet the objective, the Drought Severity Evaluation Tool (DSET) was created as a web-based drought reporting tool that:

Results

The Drought Severity Evaluation Tool (right) shows the Navajo Nation boundaries and the 85 rain-gauge locations. This map displays the CHIRPS-generated Standardized Precipitation Index (SPI) values from June to August 2018. The data underscores the variability of the SPI across the Nation, particularly during the monsoon season.

DSET allows virtual calculations of historical or current precipitation levels for any area in the region, without requiring someone to manually check a rain gauge on the ground. It combines the most accurate and longest monitoring rain gauges with the remote sensing data from satellites to provide enhanced accuracy of measurements.

Conclusion

  • DSET requires less time and resource intensive on-the-ground analytics, and compiles more local data sources to ensure that the Navajo Nation has the best available knowledge to make critical decisions about their most crucial natural resource: water.
  • As of January 2021, beta testing of DSET has occurred and further modifications have been made. There are also potential future uses for drought mitigation, streamflow diversion mapping, managing water infrastructure and restoration efforts.

References

  • McCullum et al. “Satellite-based Drought Reporting on the Navajo Nation”.
  • McCullum, Amber Jean Kuss, et al. "Satellite‐Based Drought Reporting on the Navajo Nation." JAWRA Journal of the American Water Resources Association (2021).

The Drought Severity Evaluation Tool for the Navajo Nation

Keiko Nomura, keiko@climateengine.com

Image: NASA’s Western Water Applications Office

Satellite and climate datasets in DSET

  • GPM/IMERG (V06)
  • TRMM
  • Landsat (4,5,7,8)
  • Aqua-Terra/MODIS
  • Sentinel-2
  • CHIRPS
  • gridMET
  • CFS Reanalysis
  • MERRA-2
  • SNODAS
  1. integrates precipitation data,
  2. generate drought indices,
  3. produces maps and time series analyses of drought indicators for administrative boundaries on the Nation

Image: Amber Jean McCullum, Rachel Green, Carlee McClellan

The progression of drought and climate conditions in the Navajo Nation

30 of 52

Introduction

Mangroves play a vital role in carbon sequestration and contribute significantly to coastal protection in Guyana by reducing wave action and decreasing wave energy, capturing sediments, and stabilizing shoreline substrates. Present monitoring operations depend on field observations; however, more effective methods for monitoring the full extent of Guyana's mangrove ecosystems are necessary to better monitor restoration sites, identify crucial areas of mangrove decline, and facilitate protection, among other things. The mangrove monitoring system will ultimately measure extent, biomass and health of the mangroves on an annual basis. Mangroves have distinct surface spectral reflectance properties that allows them to be easily identified from other land and marine areas as well as other vegetation. Both optical and radar images were used to classify mangrove and non-mangrove features using Google Earth Engine.

What are my Objectives?

  • Create Mangrove/Non-Mangrove Forest Classification using optical and radar images.
  • Validation of Model using Confusion Matrix
  • Create Base map of mangrove forest
  • Calculate mangrove area in hectares
  • Compare 2018 Mangrove extent to 2020 Mangrove extent
  • Ground-truth for further validation

Method

Step by Step:

  • Acquire and import both optical and radar images in GEE.
  • Pre-process optical images (Cloud Masking)
  • Filter images by date and ROI
  • Calculate spectral indices from optical images
  • Reduce image collection into one image using median image reducer
  • Display false colour composites of optical and radar images
  • Concatenate optical and radar composites
  • Use training polygon for Supervised Classification
  • Display Mangrove Extent
  • Validate model using Confusion Matrix
  • Calculate Mangrove total mangrove area using sum Reducer

Results

Conclusion

References

  • National Agricultural Research Extension Institute. (2015). Mangroves. Retrieved November 04, 2021, from NAREI: https://narei.org.gy/departments/mangroves/

A National Mangrove Monitoring System for Guyana

Kim Chan: kbagot19@gmail.com

Figure 1: Pilot area Pomeroon-Supenaam (Region 2 Guyana)

Case Study

The North-Eastern Coast of Pomeroon-Supenaam (Region 2) Guyana was used as the pilot area to develop the spatial mangrove monitoring system. It is located at latitudes 6º 55’ 00” to 7º41’0” North, and longitudes 58º28’0” to 58º52’0” West. A combination of anthropogenic activities and natural forces has significantly altered this fragile mangrove ecosystem.

Figure 2: Ground-Truthing Map showing Target Sites

Figure 3: 2020 Mangrove Extent for Pilot Area using GEE. Preliminary overall accuracy is 0.99.

Mangrove Classification using GEE

Ground-Truthing

Figure 4&5: Preliminary analysis revealed several areas of intense change in mangrove coverage. Field validation confirmed significant anthropogenic changes and minor natural changes to the mangrove environment from 2018 baseline. Mangroves were removed to facilitate large coconut farms and other agricultural activities, also for construction of buildings for both residential and commercial purposes.

Google Earth Engine was shown as a useful tool in classifying mangroves along the coast of Pomeroon Supenaam, with an overall accuracy of 0.99.

Ground truthing activity further validated the model and highlighted key areas of mangrove losses between 2018 and 2020. These changes include anthropogenic influences (agricultural activities) and natural influences (erosion).

These results will be used by NAREI and other Government Agencies to monitor existing restoration sites, identify potential areas for new restoration sites and to develop policy framework to protect the mangroves as they protect us from the sea.

Next Steps:

  • Accuracy assessment using high resolution imagery in Collect Earth Online
  • Annual Time Series of mangrove extent (2010-2020)
  • Mangrove Health
  • Biomass

Acknowledgement:

The author would like to acknowledge the follow persons for their assistance on the completion of this activity:

Dr. Marc Simard, Dr. Temitope Oyedotun, Ms. Abigail Barenblitt, Dr. Lola Fatoyinbo, Ms. Kelsey Herndon, Ms. Kene Moseley, Mr. Brian Zutta, Ms. Zola Narine

31 of 52

Damage inspection too tedious

  1. Rapid damage assessment key to plan and deploy rescue efforts and ease property insurance claims
  2. We use satellite/aerial imagery with AI to automatically identify damaged buildings

Results

Looking for two awesome ML engineers? Hire us!

  • PhDs from Stanford University, graduating June, 2022.
  • SWE/MLE/DS roles in sustainability and climate-focussed products

DamageMap: A post-wildfire damaged-buildings identifier

Krishna Rao (kkrao@stanford.edu), Marios Galanis (margalan@stanford.edu),

Xinle Yao, Yi-Lin Tsai, Jonathan Ventura, G. Andrew Fricker

Damage detection using post-fire images: a new approach

Unlike existing technologies which use both pre- and post-fire images, we use post-fire images only o decrease image sourcing challenges (see publication.)

  • Validated globally
  • Test set accuracy: 92%
  • Performance exceeds existing models (that use both pre- and post-fire images) and matches human-level performance

Extensively validated

Damage assessment for high-profile fires available on a GEE App (link)

source: abcnews

32 of 52

Introduction

Abstract:

The coverage of mobile networks in Tanzania has embarked a lot of mobile Apps use within local communities which is enhanced due to the availability of Swahili language in Android phones. WWF-Tanzania is piloting android mobile apps use for spatial data collections, monitoring, information disseminations and awareness and advocacy which is bridging the gap between different management level data visualizations, analysis and decision making by reducing paperwork and time in data collection and increasing accuracy and real-time monitoring by putting spatial data use on both online and offline mode. The mobile apps are in three different categories based on thematic areas which are terrestrial, marine and freshwater. The communities use the Apps in Village Land Forest Reserves, Beach Management Units and Water User Associations as GPS, Data Collections, Compass Awareness, Sending spatial SMS and providing real time satellite processed data like fire and weather plus administrative and landuse as base layers. Information sharing between different projects implementing partners has been efficient (ie: easy, fast and timely) and increased transparency compared to old cumbersome manual methods. Also, the apps have increased general public awareness on communities’ sustainable conservation and responsible utilization projects.

Data & Method

Most of the data are processed on Google Earth Engine using the Python API and archived onto WWF-Tanzania Monitoring, Evaluation and Learning (M&EL) Platform. The apps contains Tanzania Terrestrial, Marine and Freshwater Priority Areas spatial data which includes the Administrative boundaries & locations, National Parks, Nature & Forest Reserves, Game Reserves and Community Based Natural Resources Management(CBNRM) Areas (ie: Wildlife Management Areas - WMA, Village Land Forests Reserves - VLFR, Community Fisheries Management Areas - CFMA & Water User Associations Areas - WUA), SAGCOT, and Economic Exclusive Zone(EEZ) area.

Results

The Apps have shown to be effective tools for landuse implementations monitoring, environmental monitoring and as a data collections/communications of current/recorded/analysed spatial data/information in the WWF-Tanzania Priority Areas. These apps are potential tools to the field personnel especially the Community Based Organizations (CBO) by applying mapping in their day-to-day monitoring, evaluation and learning activities of natural resources sustainable conservations and responsible utilizations. In additions, the tools have proven to increase transparency by disseminating digital advocacy and awareness information's to the CBOs and partners/stakeholders in Swahili. (Vitini) SMS.

Conclusion

The CBOs are managing part of the conserved areas set aside by village landuse plans and under the close supervision from Village Natural Resources Committee (VNRC), both will use the apps to participate in spatial-temporal tracking, recording, reporting, awareness, advocacy and information disseminations/sharing with/to their community as part of Citizen Science in the WWF-Tanzania priority areas by enhancing community environment degradations monitoring early warning and increases the transparency, governance and accountability from all levels of government administrations (ie: Village, District, Regional and at the Ministry).

The availability of Swahili language in the Android mobile operating system has enabled the community to participate on natural resources sustainable conservations and responsible utilizations by using the smartphones equipment also the self learning is enhanced by including all training materials in digital formats so that all reference materials are available at their fingertips for reference and knowledge sharing.

References:

  • Shapiro, A. C., Nijsten, L., Schmitt, S., Tibaldeschid, P., (2015), GLOBIL: WWF'S Global Observation and Biodiversity Information Portal. 36th Symposium of ISPRS.
  • Kapos, V. et al., 2008. Calibrating conservation: new tools for measuring success. Conservation Letters, 1, pp.155–164.

ENVIRONMENTAL DEGRADATION NOTIFICATIONS

FOR CBOs AND CITIZEN SCIENTIST

Langen R. Mathew - lmathew@wwftz.org

Please replace these image placeholders with images related your work or project, or remove completely if you want to use this space for text.

The Java Programming language was used to create the Android based mobile apps which utilizes the hardware sensors of the smartphones (ie: GPS,). The prepared data was converted into GeoJSON to allow compatibility with LeafletJS as static spatial information in the apps even if there is no network coverage. The apps were designed to allow data collections. especially waypoints and tracks, the recorded data can be exported into Keyhole Makeup Language (KML) and shared. With strong network coverage (ie: Internet Speed), the apps can retrieve near-real time NASA satellite modelled data like Fire, WIND Directions, Temperatures, etc.

Figure on right: Participants from Northern Tanzania WMAs (SOKNOT-Unganisha Transboundary Landscape) who attended the MIOMBO App training-workshop at Uhuru Hostel, Moshi, Tanzania funded by Germany -BMZ-1 Project through WWF-Tanzania.

Figure above: Webmaps for providing access to spatial data retrieval, manipulation, analysis and visualization

Above Figure: One of the Dashboard which provides visualization to CBOs and Citizen Scientist for better decision making.

Below Figures: Android app Screenshots which shows the data collection, position data/Spatial SMS sending option, Compass/Search, Alert subscription and tutorials/brochure.

The CBOs (Key Users) are:

  • Village Land Forest Reserve (VLFR)
  • Community Wildlife Management Areas (WMA)
  • Community Fisheries Management Areas (CFMA)
  • Water User Association (WUA)

The Android Applications screenshots for CBOs/Citizen Scientist to Collect, Visualize, Search and Dissemination (Alert/Earlywarning)

Mobile Applications on

Google Play Store:

              • Africa Field Kit App.
              • Miombo App.
              • Jodari App.
              • Maji App.

Free and Open Source

Software (FOSS) in use:

              • Ubuntu
              • PostgreSQL-PostGIS
              • Google Earth Engine (GEE)
              • GeoServer - MapStore
              • Python APIs
              • Open Data Kit (ODK)

33 of 52

Introduction

Method

Results

Conclusions

References

Poggio, L., de Sousa, L, Batjes, N.H., Heuvelink, G.B.M., Kempen, B., Ribeiro, E., Rossiter, D. SoilGrids 2.0: producing quality-assessed soil information for the globe. SOIL, 2021.

SoilGrids 2.0: Soil Information for the Globe

Laura Poggio on behalf of SoilGrids for ISRIC, soilgrids@isric.org

Soil information is fundamental for a large range of global applications, including assessments of soil and land degradation, sustainable land management, and environmental conservation.

SoilGrids is a global product that fulfils two main goals:

1) it is a source of consistent soil information to support global modelling, and

2) it provides complementary information to support regional and national soil information products in data-poor areas.

The SoilGrids uses remote sensing to derive environmental covariates to model soil properties at the global scale. The main covariates considered are derived from a mix of optical and radar sensors:

  • morphology (e.g. elevation, landform),
  • vegetation information (e.g. NDVI and other vegetation indices),
  • climate (e.g. precipitation, land surface temperature) and
  • human factors (e.g. land use/cover).

The workflow incorporates Digital Soil mapping (DSM) practices and adapt them to the challenges of global DSM with legacy data. It builds on previous global soil properties maps, integrating up-to-date machine learning methods, the increased availability of standardised soil profile data for the world and environmental covariates . In particular:

  • quality-assessed soil profile data derived from ISRIC's World Soil Information Service (WoSIS), with expanded number and spatial distribution of observations;
  • a reproducible covariate selection procedure, relying on Recursive Feature Elimination;
  • improved cross-validation procedure, based on spatial stratification; and
  • quantification of prediction uncertainty using Quantile Regression Forest .

SoilGrids requires an intensive computational workflow, with numerous steps integrating different software. SoilGrids is entirely based on open source software, in particular: SLURM for job management, GRASS GIS for data and tiles management, and R statistical software for model fitting and statistical analysis. Predictions were computed in a high-performance computing cluster. A dynamic geographic tiling system was developed with GRASS GIS to maximise the use of memory for each job.

The maps were obtained at 250m spatial resolution using machine learning methods. Soil observations from about 240 000 locations worldwide and over 400 global environmental covariates were used as inputs.

The quantitative evaluation showed metrics in line with previous global, continental and large regions studies. The qualitative evaluation showed that coarse scale patterns are well reproduced.

  • SoilGrids v 2.0 provides global soil information for selected soil properties, including assessment and mapping of uncertainty.
  • SoilGrids250m predictions are not meant for use at a detailed scale, i.e. at the sub-national or local level, as national data providers often have access to more detailed point datasets and covariate layers for their country than SoilGrids250m can consider
  • SoilGrids v2.0 data are available on Google Earth Engine ss community contributed datasets by ISRIC - World Soil information.

34 of 52

Introduction

Overview

Over the years, the utilization satellite-derived land surface temperature has been increasingly recognized as a significant parameter in understanding the temperature changes on the land surface of the earth and gained importance for various climate applications.

This presentation presents the relationship between satellite derived LST and air temperature measurements from a weather station located in Port Area, Manila City Philippines.

The researchers used available air temperature data from 2014 to 2018, provided meteorological agency in the Philippines, and analyze its data pattern and variability. In addition, LST was derived from MODIS satellites data through Google Earth Engine (GEE).

Problem (What we want to solve)

  • Understanding this relationship allows researchers to overcome the limitation of air temperature measured by weather stations due to the extent of scale of LST data.

Method

Results

Conclusion

Conclusion

  • The relationship of air temperature and land surface temperature obtained from the location of Manila Port Area was examined.
  • This initial work allowed the researchers to understand which parameters from the available data are associated with one another.
  • The findings obtained in this study will provide insights on the trend of both air temperature and land surface temperature and be the basis of similar studies that can be done in cities in the Philippines

References

  • A. Baloloy et al., “Spatiotemporal multi-satellite biophysical data analysis of the effect of urbanization on land surface and air temperature in Baguio City, Philippines,” ISPRS - Int. Arch. Photogramm. Remote Sens. Spat. Inf. Sci., vol. XLII-4/W19, no. 4/W19, pp. 47–54, Dec. 2019..
  • K. Gallo, R. Hale, D. Tarpley, and Y. Yu, “Evaluation of the relationship between air and land surface temperature under clear- and cloudy-sky conditions,” J. Appl. Meteorol. Climatol., vol. 50, no. 3, pp. 767–775, Mar. 2011.

Temporal Analysis between Air Temperature and Land Surface Temperature in Port Area, Manila City, Philippines

Mark Angelo C. Purio, purio.mark-angelo894@mail.kyutech.jp

Problem (What we want to solve)

  • Understanding this relationship allows researchers to overcome the limitation of air temperature measured by weather stations due to the extent of scale of LST data.

Land Surface Temperature (LST)

Data Source: MOD11A1 V6 product known as the Terra Land Surface Temperature and Emissivity

Google Earth Engine (GEE): used to retrieve the LST values from the specified point where Manila Port synoptic weather station is located

Air Temperature Data

Data Source: MOD11A1 V6 product known as the Terra Land Surface Temperature and Emissivity

Google Earth Engine (GEE): used to retrieve the LST values from the specified point where Manila Port synoptic weather station is located

LST & Air Temperature Relationship Analysis

The collected data were examined and cleaned before using it as input parameters for the correlation analysis.

The respective data were observed by looking at their trend and establishing the warm and cold recorded temperatures within the year

Trend Analysis of Air Temperature and LST

  • Air temperature is relatively high from March to May and low from December to February
  • Although at different recorded value as compared to the air temperature, same observation can be seen from the graph for the land surface temperature

Relationship between Air Temperature and LST

  • Air temperature (Tmax, Tmin, and Tmean) have a significant linear correlation with Land Surface Temperature data (LST_Day & LST_Night). This means that these parameters are directly associated with each other.
  • The relative humidity shows a weak correlation with the LST data although it shows to be significant for LST_Night.

Future Work

  • Collect more data from other weather stations to provide a more comprehensive spatial and temporal analysis between air temperature and land surface temperature.

35 of 52

Introduction

  • Fixed-site monitoring does not capture local variations in air-pollution
  • Mobile monitoring, with sensors fitted on a mobile platform, provides high resolution spatial maps of on-road pollution levels
  • We conducted mobile monitoring in Bengaluru covering residential, peri-urban, and Central Business District roads
  • We conducted 27 repeat measurements over 150 unique kms in a span of 11 months
  • A Compressed Natural Gas (CNG) vehicle was used as the mobile platform carrying portable air pollution instruments
  • Parameters include PM2.5, black carbon (BC), ultra-fine particulate matter (UFPs), and meteorological parameters

Measurement of on-road air-pollution in Bengaluru, India

Meenakshi Kushwaha1, Pratyush Agrawal1, Adithi R. Upadhya1, Jonathan Gingrich2, Jai Asundi3, V. Sreekanth3, Julian D. Marshall4, Joshua S. Apte2

1 ILK Labs, Bengaluru 560046, India

2 Civil and Environmental Engineering, University of California, Berkeley, CA 94720

3 Center for Study of Science, Technology and Policy, Bengaluru 560094, India

4 Civil and Environmental Engineering, University of Washington, Seattle, WA 98195

  • Poor road conditions
  • Road closures and unplanned neighbourhoods
  • Instrument noise and calibration
  • Manual navigation

Materials and Methods

  • Collecting 1 Hz pollution data points
  • Detecting and removing of spurious data points
  • Correcting for instrument artifacts
  • Calibrating data for reference grade measurements

Results

  • Large spatial variability was observed in UFPs and BC
  • On-road BC levels were ~ 11 times higher than ambient BC levels
  • No seasonal variability was observed for on-road air-pollution
  • Major roads were the most polluted followed by arterial and residential areas.
  • Monte- Carlo analysis revealed that a minimum of 10 repeat measurements are required to get stable spatial map of air pollution

BC

Specific study roads are in green.

Study neighborhoods are color-filled shapes.

  • Gridding to 30 m spatial resolution
  • Aggregating to daily and study period averages
  • Monte-carlo subsampling for stability analysis

Challenges

BC(μg/m3)

UFPs concentration on different road types

BC concentration on different road types

Monte-Carlo analysis

Spatial map of BC and UFPs concentration in a residential neighbourhood

UFPs (#/cm3)

Instruments used on the mobile platform

36 of 52

Who

•Undergraduate students taking General Education Geography Courses including Physical Geography and Will the World Provide?, a sustainability course

•~ 80 students per year

•Typical student majors include Sports Management, Education, Exercise Science, and Geography

How

What

•Google Tours to explore different environments

•Google Tours to explore different surface features

•Create their own Google Tours as a final project

•Timelapse to explore changes to climate and environment

•Timelapse to explore human-driven changes to Earth

Why

Social Impact: Undergraduate Education

Melinda Shimizu, Ph.D. melinda.shimizu@cortland.edu

•Dynamic views and photos in the Google Earth interface are far better at capturing and communicating Earth features than static photos

•It is interactive and students thrive when presented with interactive opportunities to learn

•The Timelapse feature is so helpful to showing changes to environments in an accessible way

37 of 52

Why mangrove phenology is important?

Estimating phenological metrics

So far..

Next..

References

  • Kamruzzaman et a. (2017) Forest Ecosystems, 4: 16. DOI: 10.1186/s40663-017-0104-0
  • Mandal et al. (2020) Tropics, 29 (2) 41-55. DOI: DOI: 10.3759/tropics.MS19-11

Elucidating mangrove forest phenology in the Sundarbans

Mohammad Shamim Hasan Mandal and others

Postdoctoral Researcher, Graduate School for Advanced Science and Engineering Hiroshima University

mshmandal@gmail.com | mandal@hiroshima-u.ac.jp

Phenological metrics::

  • start of the season (SOS)
  • time of maximum greenness (MaxGreen)
  • end of the season (EOS)
  • length of the season (LOS)

Pre-processing and

Time series in

Google Earth Engine (GEE)

Cloud masking and filtering

Phenological parameters

Spectrum using FFT

Smoothing

MODIS (MOD13Q1) 16-day

Remote Sensing Data

NDVI, EVI

Time series: four forest types

Pneumatophore Vivipary

Sundarbans mangrove forest

  • Mangrove forests are the most carbon-rich ecosystems.
  • Phenology provides measurable links to study climate change

But,

  • Its study, need long-term observation
  • Baseline information on phenology is poor

F1 - Heritiera fomes dominated,

F2 - Excoecaria agallocha dominated,

F3 - Ceriops decandra dominated,

F4 - mixed forests of F1, F2, and F3

A clear annual seasonality in canopy greenness

Phenological calendar

Future study

  • MODIS 8-day MODIS surface reflectance data
  • Indices such as gNDVI, NDMI, EVI2 should be tested

Lesson learned

  • MOD13Q1 product is 16-day interval
  • Both NDVI and EVI can be problematic in high biomass area:

38 of 52

Introduction

Visualize the Environmental impact of Project Finance

My aim is to demonstrate the environmental impact in the implementation of new projects, through the use of Google Earth, in order to allow the various authorisation levels to make decisions that include both saving costs and the environment concern.

Bullet List

  • To create a sample Earth project that others could learn from
  • To use all the new features of creation tools in my story
  • To document the step-by-step process so that new learners could re-create it!

Method

I created a project in Google Earth, detailed with pictures and plans of the project including the necessary information to describe the issue.

In order to represent the situation in the best possible way, and to allow the boards to understand the problem, I shared the project so that each individual could analyse the issue in the best possible way.

Bullet List

  • Updated the project drawing
  • Share the content of the project
  • Interact with C-suits in order to demostrate hte anaylsis and theencironmental impact of the project

Results

Changes to the initial plan to create gas lines affecting the turtle reserve were accepted and a new project was considered.

the real difficulty was to instil within an old-style managerial class the possibility of assessing the feasibility of new projects through the use of open source resources such as Google Earth

Conclusion

Bullet List

  • make it clear that just because it is a new instrument does not mean that it is not as reliable as the old methodologies
  • environmental impact is important in the evaluation of financial projects
  • Google Earth can make a difference in the decision making process

References

  • Nicolò Spelgatti - Visualize the Environmental impact of Project Finance

Visualize the Environmental impact of Project Finance

Nicolò Spelgatti, nic.spelgatti@gmail.com

Logo

Please replace these image placeholders with images related your work or project, or remove completely if you want to use this space for text.

Please replace these image placeholders with images related your work or project, or remove completely if you want to use this space for text.

Please replace these image placeholders with images related your work or project, or remove completely if you want to use this space for text.

Please replace these image placeholders with images related your work or project, or remove completely if you want to use this space for text.

39 of 52

Introduction

Spatiotemporal Assessment of Forest Cover Change

Understanding changes of forest cover is important for identifying hot-spot areas and drivers of change to form a basis for formulating effective conservation and management strategies. However, study using a combined method of phenology-based threshold classification (PBTC) and Google Earth Engine (GEE) is still unavailable in Sri Lanka and in South Asia. Here, we the GEE along with the Landsat TM and Landsat OLI collections and applied PBTC method to map the spatial distribution and analyze temporal changes of forest in Vavuniya District, Northern Province, Sri Lanka since 2001.

Our study aims:

  • To modify enhanced vegetation index (EVI) threshold values for dry monsoon and open forest in GEE for Landsat 5 TM and Landsat 8 OLI
  • To map the spatiotemporal distribution of forest cover using Landsat 5 TM and Landsat 8 OLI images and PBTC method in GEE.
  • To assess the forest trend (gain and loss) and annual rate of changes between 2001 and 2020 in Vavuniya District, Sri Lanka.

Methodology

We acquired and processed 234 collections of Landsat 5 TM and Landsat 8 OLI during dry-phenology season (May to September) to map the land covers for 2001, 2006, 2010, 2016, and 2020 to obtain the images for analysis (Figure 2).

We applied the minimum and maximum EVI threshold values for dry monsoon forest and open forest.

EVI thresholds as listed below:

  • Landsat 5 TM: Dry monsoon forest (0.5027 – 0.776), and Open forest (0.4845 – 0.501)
  • Landsat 8 OLI: Dry monsoon forest (0.650 – 0.882), and Open forest (0.631 – 0.648)

Accordingly, land cover classification and spatiotemporal mapping is performed by executing PBTC method in GEE. Our study yielded five PBTC land cover maps.

We assessed the rates of annual changes for whole district’s forest and its four Divisional Secretariats divisions.

Conclusion

Assessment of Forest Cover Change using the Phenology-based Threshold Classification and Google Earth Engine

Sharaniya Vijitharan (st121566@ait.ac.th), Manjunatha Venkatappa, Nophea Sasaki

Figure 1. Visualization of forest cover types

Naturally-regenerating forest types

Study Site

Methods

Figure 3. PBTC in GEE

Figure 2. Methodological framework

Results

Conclusion

Our research achieved the average overall accuracy of 87% and mean Kappa coefficient of 0.83.

Forest area expanded between 2001 and 2010, while a reduction measured during 2010-2020.

We assessed:

In 2020, 57.5% of district’s area was covered by forest.

  • Dominant forest type was dry monsoon forest with 52.2%.
  • 5.34% of total land area occupied by open forest.

Forest cover changes in DS divisions:

Vavuniya – increased until 2020 and then declined

Vavuniya South -- slightly increased

Vengalacheddikulam – gradually expanded

Vavuniya North – continuously declined by 19.2% (2001-2020)

Figure 4. Overall trend of forest cover

Figure 5. Forest cover change in DS divisions

Forest cover changes between 2001 and 2020:

  • Total district's forest: -0.09%
  • Dry monsoon forest: -0.3%
  • Open forest: 3.62%

From 2010 to 2020, total forest cover reduced by 6.04%. Forest area expanded between 2001 and 2010, while a reduction observed during 2010-2020. Dry monsoon forest cover declined continuously for 19 years. Overall open forest cover was slightly increased.

Highest forest cover was in Vavuniyabut nut it declined from 83.8% in 2001 to 67.7% in 2020.

Classification of open forest is challenging due to its smaller extent mixed with croplands, and soil reflectivity during dry season.

Without GEE, it would not have been possible to access data since 2001 as such data are rarely available in Sri Lanka.

Figure 6. Land cover maps of Vavuniya District, Sri Lanka

Main Points

40 of 52

Introduction

Climate change and human activities have had a significant impact on the forest in recent decades. Forest fires are caused by climate and human usage, resulting in infrastructural damage and loss of life. Several studies learn more about the adverse effects of forest fires on flora, biomass, and soil characteristics (Daz-Delgado et al., 2002; Verma et al., 2019). Forest fires in Nepal's forests are quickly rising each year. According to studies, more than 90% of vegetation fires in Nepal are likely caused by humans, with forest fires affecting 3 to 5 million hectares yearly (Durbar & Kathmandu, 2018; Yasmi et al., 2019). Since 2005, Nepal has lost an estimated 200,000 hectares of forest cover per year due to forest fires, according to the report (Chaudhary et al., 2016; Paudel, n.d.). The forest fire trigger by a prolonged drought and lack of rain. Its smoke pollutes the air and forms a thick layer in the atmosphere. As a result, forest fire identification and monitoring research are necessary to determine its kind, magnitude, and impact and develop a risk map and early warning system to reduce its repercussions.

Objectives

  • To reduce forest fire, people should be aware and inform of its long-term effects on ecosystems
  • Using such an approach can reduce the risk of human fire and natural fires, mapping hotspot areas with high and moderate burn severity.
  • In addition, it guides to plan and preparedness time for the local government and community

Method

In Gandaki Province's trans-Himalayan region, Manang has dense pine forest and approximately 25% of the Annapurna protection area (Bhattarai & Conway, 2021; Huddart & Stott, 2020). Currently, there is no road access in the district.A wildfire raged for two months in Thanchok and Kote's woods in January, destroying approximately 700 hectares of forest cover (Wildfire Burns 700 Hectares of Forest Forest Cover in Manang, n.d.; Wildfires Destroy Hundreds of Hectares of Forest in Nepal - Xinhua | English.News.Cn, n.d.). It contains a highly flammable thick pine forest, which may have been ignited by accident or on purpose.

The research area is estimated using the Landsat-8 satellite image indexes, object-based categorization, fire area, and burning intensity. It uses Google Earth Engine to generate a potential forest fire map of Manang generated using different sources for the forest fire area’s burn severity.

  • Forest fire area detection using Landsat eight satellite images through its vegetation indices has emerged as a possible way to estimate wildfire.
  • It's the first time such an approach assesses fire risk across an entire Himalayan region, and the methodology still has certain limitations.
  • It identifies possible high, moderate, and low burn areas for reducing risk and preparedness.

Results

The differenced Normalized Difference Vegetation Index (dNDVI) and differenced Normalized Burn Ratio (dNBR) map of the study area is the difference between the pre-fire and post-fire NDVI and NBR obtained from the satellite images to estimate the burn severity. dNDVI index obtained by subtracting the NDVI index after the fire from the NDVI index before the fire. Higher values indicate dense vegetation cover and vice versa. A higher value of dNBR indicates more severe damage, while areas with negative dNBR values may indicate regrowth following a fire. The differenced Green Normalized Difference Vegetation Index (dGNDVI) map for distinguishing stressful vegetation and the Relative differenced Normalized Burn Ratio (RdNBR) map measures the spectral variation.

Conclusion

  • calculate bands and vegetation indicators, such as NBR, NDVI, dNBR, dNDVI, RBR, RdNBR, GDVI dGDVI, and BAI, to identify the geographical extent of the fire.
  • It shows burnt areas and burns intensity in the Manang district using wavelengths near-infrared (NIR) and shortwave-infrared (SWIR).
  • Forest fire severity is divide into eight categories based on the severity of the burn, with about 26% falling into four categories: low, moderate, high, and higher severity. Furthermore, around 30% of unburned and low-severity fires are estimated to be about 37%.
  • Thus, It indicates that approximately equal sections of the land remain unburned and at risk of fire.

References

  • Pawan, T. (2020). Evaluating a Fit-For-Purpose Integrated Service Platform for Climate Change and Land Information of Rural Mountain Community. J Remote Sens GIS, 9, 278.
  • Singh, R. (Ed.). (2021). Re-envisioning Remote Sensing Applications: Perspectives from Developing Countries. CRC Press.
  • Thapa, P. (2021). Using Geospatial Technologies for Disaster Management in Developing Countries..

Detecting Forest Fire using Landsat Images: A Case Study of Manang District, Nepal

Pawan Thapa, pawan.thapa@ku.edu.np

41 of 52

What and why?

Watch my Wilderness

The Wilderness Society (Australia) is an ENGO that campaigns to protect Australia’s iconic wild places and biodiversity. We empower the community to influence the decisions about the places they love. One of the tools we are developing is a community powered deforestation monitoring platform. The Watch my Wilderness (WmW) platform provides easy access to Sentinel-2 data to compare images to detect land clearing/deforestation. The users can trace suspected clearing and submit it to our database for verification. We aim to contribute to open source training data for change detection models that are being built to assist with biodiversity conservation, vegetation management/compliance and carbon related projects.

WmW aims to:

  • Empower our community to monitor land use change across Australia in an easy to use platform
  • Create a land clearing database that can be used as training data for better auto-change detection of Australia’s diverse landscapes
  • Detect clearing events in near real time and expose the industries behind the events
  • Raise awareness of Australia’s deforestation issue

How?

Watch my Wilderness uses Google Earth Engine, the Google Cloud platform, and Google Maps to:

  • retrieve processed Sentinel-2 imagery (both natural and false colour)
  • run our periodic scripts for auto change detection and pre-rendered cloudless mosaics
  • provide long-term bucket storage of script outputs that are then accessed by the WmW application and used as map overlays
  • embed the data and provide an interface for interacting with the map components (eg tracing the clearing events, location searches etc)

Did it work?

We have released a restricted access version of Watch my Wilderness to test user needs for monitoring Australia. Sentinel-2 EOM provides high frequency, medium resolution imagery and Google makes it easily and freely available for grass roots platforms like Watch my Wilderness. The users found the platform to be very useful for remotely monitoring land without having to go into the field. Users also found it useful to have access to GIS professionals to assist in image interpretation and to verify clearing. All users wanted more advanced auto detection to guide the monitoring, which the WIlderness Society is currently working on. The platform is used by a diverse group of people, mostly non-GIS professionals, and has a verified clearing database of over 400 clearing events from 10 trial users in Queensland alone.

What next?

  • As a community-based campaigning organisation, we knew that our deforestation campaign would benefit from the help of our community, as well as provide a platform for our allies to use for their own monitoring programs. This led us down the citizen science / Google Earth Engine path and WmW.
  • The Wilderness Society has developed a WmW prototype and is currently developing a better platform that will go live in 2022, based on the basic principles of the prototype, but with enhancements to the base code and the user interface/experience
  • We will use the WmW clearing detections to train our auto-detection algorithm so that manual verifying is made easier
  • We would like to provide this data to other training databases

References

Land clearing and deforestation monitoring in Australia

Rachel Fletcher, rachel.fletcher@wilderness.org.au

Citizen science:

Australia has a high rate of deforestation and land clearing across a large area. The Wilderness Society is harnessing the power of our community of volunteer citizen scientists to detect and record clearing events across Australia.

42 of 52

43 of 52

Introduction

Sustainable Supply Chain

Developing a sustainable supply chain from seed to delivery and everything in between is something I’m passionate about. I’m meeting with people responsible for these initiative at my clients who mostly reside in the SLC industry.

  • Lead with the Alphabet ecosystem of tools.
  • Earth Engine, Maps, Fleet routing optimization, X Chorus, Waymo, Waze, X Wildfires, Tree Top and others.

Method

  • IShow powerful case studies where we’ve made a positive impact on enterprise carbon footprint while not increasing the overall cost.
  • Create demos and architecture to lead the way.

Results

  • My goal is to help my accounts get to carbon neutrality. This is a story I’m just starting but I’m looking forward to the journey and results.

Conclusion

  • There is no tomorrow on improving the health of our planet and lowering global temperatures. Be it emissions, cement manufacturing, deforestation, sustainable supply chain or any other eco issue the time is now or never. .
  • The logo is my children. I want to leave them a better world.

References

  • Ullamcorper eget sem. Interdum et malesuada fames ac ante ipsum primis in faucibus.
  • Nullam aliquam nisl risus, nec pretium erat vehicula a Maecenas duiex, euismod sit

Sustainable Supply Chain & Logistics

Scott Hitchcock, scotthitchcock@google.com

44 of 52

Introduction

Google Earth Election Maps

Leverage the power of open data and free Google apps to educate the public about elections.

Visualization of election data provides a clearer interpretation of election outcomes and uncovers patterns which may be used to measure performance.

Oblige Google Maps Election Project is an online workshop and toolkit that demonstrates how to transform election data into useful and understandable maps that improve understanding of elections. Oblige demonstrates how to easily leverage free Google Apps for election visualization, analysis, strategy, and public education.

Dissipation of local news outlets has created an aperture in the media marketplace for the visualization of local election data. News media data visualization resources are strained across markets. The election mapping market is dominated by proprietary mapping professionals and hobbyists using softwares with steep costs and learning curves. Very few resources provide interactive maps for free public consumption or iterations.

Oblige Google Earth Election Maps Mission

  • Instill more public trust in the democratic process.
  • Communicate election results in a clear understandable and open medium.
  • Educate the public how to analyze and interpret election results through interactive and open mediums.

Approach

Clean data may be mapped in Google Earth manually and methodically, creating interoperable maps that can be shared and further analyzed. Building upon the principles of the Open Election Data Initiative developed by Google, NDI, and the USAID; this project seeks to demonstrate how to visualize that data for better comprehension and wider consumption.

Method

  • Acquire, clean, verify, and format election data from supervisors of elections or government agencies following Open Election Data Initiative principles.
  • Acquire and verify GIS shapefiles of voter precincts from government agencies and import to Earth, modify or draw your own.
  • Map results of analysis in Google Earth and upload to My Maps for interactive shareability.

Insights

The 2020 Presidential Turnout Map to the right demonstrates an in depth analysis of election results. Election and voter data was gathered and cleaned using Google Sheets and Drive. Calculations were mapped in Google Earth and saved and shared to Google Maps and disseminated. The results shown in the map present a clear pattern that economically depressed areas had lower turnout in the election.

The ability of this data to be easily read and shared interactively on mobile devices in the field allows political campaigns and organizations to allocate resources nimbly and strategically to effectuate greater election participation.

How may we Oblige?

Fostering a better understanding of community through elections data visualization using Google Earth to engender more trust and understanding of democratic elections.

Leverage free tools to educate and improve performance outcomes and identify disparities.

References

Oblige: Google Election Maps

Seth Platt, Sethplatt@oblige.org

Barriers are access to election data/clean data under non-transparent or unorganized governments.

Resources needed to map election results: Human, GE processing power, data.

  • Sethplattsmaps.com
  • Oblige.org

The map above delineates Democrats and Republicans elected as precinct committee people per voter district in Broward County, Florida. This map identifies areas of strength and weaknesses in the organization's’ County wide human resources.

Publicly available addresses of elected members were uploaded to Google Earth from Sheets and demarcated manually in Google Earth according to how many resided in each voter precinct. This map define a resource allocation strategies.

45 of 52

States with FAW presence

Introduction

Fall armyworm (FAW) arrival in India

  • FAW has been a big menace for Indian farmers since its arrival in July, 2018.
  • The maize farmers specially has suffered huge yield loss.
  • Understanding of FAW behaviour is crucial for its management.
  • With help of Digital tools and platforms like Google Earth Engine and Machine learning tools, we understand pests’ pattern and predict its risk early on in season to manage it appropriately

Goal

  • To develop a machine learning model to forecast and predict FAW risk early on in season
  • To develop digital platform for farmers in collaboration with Corteva Agrisciences’ field team to provide early alerts against the pest

Method

Results

Conclusion

References

Google Earth Engine Tutorial

Wu, Q.-L.; He, L.-M.; Shen, X.-J.; Jiang, Y.-Y.; Liu, J.; Hu, G.; Wu, K.-M. Estimation of the Potential Infestation Area of Newly-invaded Fall Armyworm Spodoptera Frugiperda in the Yangtze River Valley of China. Insects 201910, 298.

Forecasting Fall armyworm using multi-sensor data for maize farmers in India

Sonal Bakiwala, Sravani Gunda, Parmita Ghosh, Raman Babu, Anu Swatantran

Farmer

Agronomist

IBM Weather

Data

AWS

Database

ML Model

FAW Risk Probability

Prediction

Satellite Data

Engineered Features

Data Digitization from Farmer location

Step by Step

  • Ground truth data like location, severity levels of damage by pest, plant age was collected by farmer survey
  • Using Google Earth engine pro, data was digitized, and locations were converted to polygons format
  • The Weather data was collected from IBM weather API
  • The Sentinel-1 SAR images were acquired using Google Earth Engine Platform
  • The Corn Map was generated with sentinel-1 data by building SVM model with 92.1% accuracy
  • Using the corn map and integration of multi-sensor data, FAW risk model was developed with 88% accuracy

Haversine’s law

Distance and Bearing

Sentinel-1 SAR

Catboost Model Test Accuracy 88%

Total

High

Medium

Low

Healthy

8197

1532

2994

2896

775

Ground Truth data for model building and validations

Ground truth data for FAW incidence ground survey

Predicted FAW Severities from multi sensor model(14 days in advance)

Misclassified

severity

Challenge

We did not have information on farmers if they applied any chemistry to manage FAW. Its possible predictions are correct, but farmers may have applied solutions and they don’t witness heavy damage due to FAW

FAW has been a menace for the entire country since 2019. Extensive research and studies have been done to manage FAW better through multi-disciplinary sciences. We have produced a strategy to exploit multi-sensor data and integrate them together to forecast FAW risk better. We leverage information and data from farmers directly. We understand that its very important to help the community at large especially small farmer holders. Using high resolution satellite data and weather data we can deliver early alerts 14 days in advance for FAW risk and help farmers (especially small holders) to mitigate the FAW risk and not suffer yield loss because of FAW damage. More validations across country and cross seasonal validations are required.

Conclusion

High Severity

Medium Severity

Low Severity

Corn areas

Ground truth

Cross validation

Accuracy – 92.1%

Corn area Prediction using Sentinel-1

---Internal Use---

46 of 52

Introduction

Increase the resilience of cities

Nowadays our cities are flooding with increased frequency, due to more severe weather events but also to increasing anthropogenic pressure (e.g. soil sealing, land subsidence, uncontrolled development, etc.).

It is therefore necessary to increase the resilience of our cities and facilitate their adaptation to current and future extreme flooding events

What about the challenges?

We want to equip communities with the appropriate flood risk intelligence for supporting preparedness and resilience:

  • assess and forecast urban flooding;
  • support multiple users in designing and assessing multiple mitigation measures in urban areas and contribute to increasing awareness and capacity building for supporting flood climate resilience among different stakeholders;
  • exploit Open and Big Geospatial (Google Earth Engine, OSM) and Climate Data (Copernicus) and AI-based flood hazard and damage models.

Method

SaferPLACES is a cloud-based climate service for assessing pluvial, fluvial, and coastal flood hazards and risks under current and future climate scenarios. A DIGITAL TWIN of the urban environment can be generated in few steps in any city of interest, to provide accurate and actionable insights to flood risk assessment and mitigation.

The algorithms

  • Safer_RAIN is a DEM–based filling-and-spiling hazard model exploiting high-resolution LiDAR DEMs available for urban areas; it requires few input data and can generate flood hazard maps in real-time for several rainfall events a accounting for infiltration (see e.g., Samela et al., 2020);

  • Safer_RIVER is DEM-based flood-hazard model developed for fluvial flooding; a fast-processing binary classifier uses a geomorphic flood index for delineating flood-prone areas (see e.g., Tavares da Costa, et al., 2019);
  • Safer_COAST is a dedicated zero-dimensional DEM-based flood model characterizes coastal flooding hazards. It uses a fast-processing Region Growing Algorithm for coastal flood mapping;
  • Safer_DAMAGE is a Bayesian data-driven multivariate damage assessment model for predicting direct flood losses in residential areas (Paprotny et al., 2020).

Results

SaferPLACES platform provides a cost-effective and user-friendly cloud-based solution for flood hazard and risk mapping. Moreover SaferPLACES supports multiple stakeholders in designing and assessing multiple mitigation measures such as flood barriers, water tanks, green-blue based solutions and building specific damage mitigation actions.

Conclusion

SaferPLACES capabilities

  • employ innovative climate, hydrological and raster-based flood hazard and economic modelling techniques to assess pluvial, fluvial and coastal flood hazards and risks in urban environments under current and future climate scenarios;
  • the platform provides a cost-effective and user-friendly cloud-based solution for flood hazard and risk mapping;
  • supports multiple stakeholders in designing and assessing multiple mitigation measures such as flood barriers, water tanks, green-blue based solutions and building specific damage mitigation actions.

References

  • Stefano Bagli (Gecosistema s.r.l., Italy): stefano.bagli@gecosistema.it
  • Attilio Castellarin (Univ. Bologna, Italy): attilio.castellarin@unibo.it
  • Federica Bonaiuti (Univ. Bologna, Italy): federica.bonaiuti3@unibo.it

SaferPLACES Global Platform: A Digital Twin Solution for

Flood Risk Intelligence in Urban Areas

Stefano Bagli, stefano.bagli@gecosistema.com

47 of 52

Introduction

Satellite data to monitor montane forest recovery

  • Many tropical montane forests are degraded by fire, land use and climate change
  • Little is known on how these forests recover following degradation, and where active and passive restoration operate
  • In this UN Decade on restoration it is crucial to elucidate on ecosystem recovery in highly biodiverse systems like tropical montane forests
  • We use 20-year time series of multispectral Landsat data to track recovery trajectories of degraded montane forests across the tropics

Method

Results

Recovery trajectories vary after different kinds of disturbance. Large areas of burned forests show no recovery at all, while other areas recover. Disrupted recovery and no recovery areas can be targeted for assisted and active restoration interventions.

Conclusion

Challenges

  • Free satellite data sets of very high resolution multispectral data (<10m spatial resolution) covering time spans of several decades are not freely available. Hence, Landsat sensors are the only option for long-term recovery assessments
  • Memory issues limit pantropical image processing and recovery classification of 30m resolution data during workflows in GEE.

Assessing recovery trajectories in disturbed tropical montane forests through multi-temporal satellite data

Tina Christmann and Dr Imma Oliveras, tina.christmann@worc.ox.ac.uk

A promising tool for restoration planning

Satellite-based recovery classifications can guide planning and decision making on ecosystem restoration in inaccessible ecosystems such as tropical mountains. For instance, by identifying areas of successful and unsuccessful recovery, restoration priority hotspots can be scoped and suitable restoration methods chosen. Such methods can als be transferred to monitor existing large scale active restoration projects (e.g. reforestation projects) and improve transparency and accountability of ecosystem restoration.

Fig. 1: A tropical montane forest degraded by fire in the Brazilian Atlantic Forest (own photograph)

Fig. 2: Workflow of degradation and recovery classification

Fig. 4: Recovery trajectories in different degradation areas

Fig. 3: Spatially explicit classification map for four recovery classes based on trend analysis

Degradation and recovery classification

  • First, we identify tropical montane forests with an elevation and land use mask. Then we identify degradation and potential areas of recovery using Hansen Forest Loss and Modis Burned area datasets.
  • Second, using trends analysis, we categorize recovery trajectories in potential areas of recovery and calculate recovery ‘budgets’ for different geographic areas.

48 of 52

Introduction

Story begins

Having previous work experience with poverty mapping and being locked at home during pandemics for months, I have decided to spend my time taking care of my newborn baby and experimenting with new AI methods. Not only to implement, however also to to tune it all and possibly check for freely available input data. After analysis, I focused on two research papers. “Classifier” and “Detector”.

Goooooaaaaal

  • combine 2 existing poverty mapping methods and tune with own ideas on freely available data/SW/HW
  • learn by practising
  • write a paper, if project is successful

How it was done

In Google Earth, I manually located polygons of the target country, where very high resolution (0.3m/px) is available and downloaded them. Due to resolution and HW restrictions, only cca 200 out of 1500 administration regions were used in this project.

In Earth Engine, Sentinel 2, VIIRS nightlight imageries and Worldpop datasets were aggregated and downloaded for AOI.

Virtual Machine with GPU from Google Colab was used for the training and inference phases of the modelling.

Fastai2 and Pytorch provided machine learning frameworks for the creation of classifiers(arch ResNet34) and detectors(arch Yolov5).

Linear models for final poverty mapping and validation were implemented based on the scikit-learn library.

Results

Creating “classifier” and “detector” separately, I was able to reach comparable results with the original papers. Furthermore, the resulting combined ensemble model gained high poverty prediction accuracy (r_squared=0.82) by exploiting existing methods, different model architectures and satellite image resolutions. For more detailed description please see published paper.

Conclusion

  • With all of the AI/SW/cloud computing advancement nowadays, individuals having enough patience and coding skills, are able to implement latest state-of the-art algorithms, which used to be a domain of research teams and corporations.
  • Big thanks to all people and organizations for sharing the know-how and making the research and tools available for usage in projects for public good
  • Experiment, Fail, Learn, Repeat :)

References

  • Yolov5 detector
  • Fastai

Poverty mapping from space

Tomas Sako, tomassako@gmail.com

Yolov5 framework from Ultralytics with sample detections, Google Earth image (2016 CNES/Airbus)

VIIRS Nighttime Composites Version 1

49 of 52

Composites derived from Landsat-ETM+ 30m (left) and WorldView2 2m resolution data (right).

Introduction

NSF South Big Data Hub’s

3D Wetlands Project

  • Goal: Improve data quality & resolution to aid in assessment of:
    • flood risk
    • conservation of sensitive habitats.
  • Geographic scope: Gulf of Mexico’s coastal floodplain from Texas to Florida.

Produced Datasets

Datasets in GEE

DEM & Habitat Classification Maps are now on GEE

  • See maps
  • Create & analyze transects
  • Summarize class % coverage
  • Download data for further analysis

github.com/luislizcano/3D-wetlands-app

  • Build SDMs using ground truth
    • (mangrove+seagrass)
  • Improved classification algorithm
  • Convolution to classify/smooth using texture
  • Habitat vulnerability to sea level rise

Next Steps

3D Wetlands : 2m Habitat Class + Elevation in the SE USA

Luis Lizcano-Sandoval {luislizcanos@usf.edu}, Tylar Murray {mail@tylar.info}

Target area WV2 coverage

DataSet 1 : Digital Elevation Model

  • Manual merging of many airborne LiDAR datasets.
  • Estimates ground level from LiDAR point clouds.

DataSet 2 : Land Cover

  • Automated land cover classification on multispectral commercial satellite imagery.
  • Selected classes & created decision tree.
  • Ingested TBs of data via CAD4NASA.

50 of 52

The Problem

Science informs management

The Nature Conservancy (TNC) and partners are working to reveal and understand coral health at different scales to answer a range of vital questions for resource managers, governments, and conservationists. Scientists are collecting data using different technologies and platforms, each of which provides an important layer of information to inform marine conservation and management initiatives. Airborne imagery from the Arizona State University's Global Airborne Observatory (GAO) can be used to derive spatial data such as live coral cover and habitat complexity, or rugosity. These maps will help coral managers identify specific locations to outplant coral to maximize survival rates.

In the past, site selection for coral outplanting was based on local knowledge and observations, with little quantitative information involved in this process. However, in November 2019, The Nature Conservancy and Arizona State University’s Global Airborne Observatory took a step forward by collecting data and developing a new tool to inform site selection. Coupled with local knowledge of ideal in-water conditions, this data allowed for selection of sites that improved coral survival.

The Data

Arizona State University's Global Airborne Observatory (GAO) is an airborne laboratory that collects high resolution (1m) coral health and habitat information to inform effective coral outplanting site selection with increased survivorship. The GAO plane flies at an altitude of 1000m and collects 427 samples of light per pixel that correspond to the chemistry of corals, generating 324 GB/hour of data. Data is processed on a cluster machine matching spectrometer information with high-precision positioning data at an accuracy of 7-10cm.

Hyperspectral imagery was collected by ASU's Global Airborne Observatory (GAO) for The Nature Conservancy in Saint Croix, U.S. Virgin Islands and the Dominican Republic. ASU derived high-resolution maps from this aerial data representing: bathymetry, live coral cover, fine habitat complexity, algal cover, seagrass cover, and sand cover. TNC scientists used water surface drone imagery to collect high-precision GPS data and to predict percent live coral to train the machine learning model. These maps are made available here for use by resource managers and conservation scientists.

The GEE App

On the Ground Results

The Coral Outplanting Siting Guide app allows on-the-ground coral managers to find sites that meet suitable criteria ranges of different variables such as percent algal cover, habitat complexity (rugosity), live coral cover, and bathymetry. This scientific method for selecting sites can guide conservation and management efforts when coupled with local knowledge. This app is available at CaribbeanMarineMaps.tnc.org and a similar app is currently under development for the Hawaiian islands by The Nature Conservancy and partners.

Reference:

  • Schill SR, Asner GP, McNulty VP, Pollock FJ, Croquer A, Vaughn NR, Escovar-Fadul X, Raber G and Shaver E (2021) Site Selection for Coral Reef Restoration Using Airborne Imaging Spectroscopy. Front. Mar. Sci. 8:698004. doi: 10.3389/fmars.2021.698004

Coral Outplanting Siting Guide

Valerie Pietsch McNulty, valerie.mcnulty@tnc.org

© Paul Selvaggio

© Paul Selvaggio

© Paul Selvaggio

This tool was successfully implemented during a coral mania organized by Fundación Grupo Punta Cana in Cabeza de Toro, Bavaro. This event brought together dozens of volunteers belonging to NGOs, Dominican environmental authorities, dive operators and other local stakeholders aimed to outplant 1,711 Acropora cervicornis fragments into different sites. By October 2020 (11 months after outplanting), survivorship remained above 76%. These results demonstrate higher than average success rates for coral outplant survival for this species.

51 of 52

Introduction

Globally, about 930 million hectares of forests are classified as degraded forests, while another 10 Mha are deforested annually. Foreseeing this danger, the New York Declaration on Forests and the UN Decade on Ecosystem Restoration set the target to restore 350-500 Mha of the degraded ecosystems by 2030. At COP26 of the UNFCCC, world leaders agreed to completely stop deforestation by 2030. To implement these agreements at scale, identification of the appropriate degraded and/or deforested lands is prerequisite. However, little study exists on this important identification. Such identification at scale needs scalable technology along with high computing power.

What was our aim?

  • To develop a novel approach for identifying potential degraded forests for restoration (PDFR) and prioritizing appropriate restoration strategies.
  • To estimate the potential carbon sequestration that might be realized from the restored forests between 2021 and 2030.

Method

  • We developed a PDFR approach by applying phenology-based threshold classification (PBTC) on forest cover data at a 30-m resolution in Siem Reap Province, Cambodia, using Landsat data in the Google Earth Engine (Fig.4 and Fig.5)
  • We determined the appropriate areas for restoration were evergreen forest, semi-evergreen forest, deciduous forest, and flooded forest.
  • With the PDFR, we prioritized forest restoration according to the following degradation levels: critically degraded, highly degraded, moderately degraded, and slightly degraded forests.
  • We estimated the potential carbon sequestration that might be realized from the restored forests between 2021 and 2030

Results

We were able to prioritize the areas for forest restoration according to the following degraded levels: critically degraded (~96,693 ha), highly degraded (48,878 ha), moderately degraded (46,487 ha), and slightly degraded forests (75,567 ha).

If all these degraded forests were restored, 193.73 TgCO2 could be sequestered in Siem Reap between 2021 and 2030 (Fig. 6).

Depending on carbon prices, US$0.8 – 26.6 billion could be generated.

Conclusion

Main points

  • Our novel PDFR approach on GEE makes it possible to identify the degraded forests in the tropics at scale.
  • The PDFR approach could become a useful tool to assist the large-scale forest restoration planning on automation.
  • The PDFR approach may also be used to facilitate the monitoring, reporting, and verifying activities as required under the REDD+ scheme of the UNFCCC.

References

  • Venkatappa et al. (2019). https://doi.org/10.3390/rs11131514
  • Venkatappa, M., Sasaki, N., Stefan, O., Soben, K., Benjamin, S., 2021. Identification of the Potential Degraded Forests for Restoration in the Tropics –Implications for Carbon Sequestration and Revenues. Journal of Cleaner Production. (Under review)

Identification of the Degraded Lands for Forest Restoration on Automation

Manjunatha Venkatappa, Nophea Sasaki manjunathagis@gmail.com | www.leetintel.com | www.ait.ac.th

Figure 1. UN Ecosystem Restoration by 2021-2030

Figure 2. Tropical forest degradation levels

Figure 3. Level of forest degradation

Figure 4. Forest transition and level of forest degradation

Figure 5. Forest cover change between 1990 and 2018

Figure 6. Potential carbon sequestration (million tons) between 2021 and 2030 and benefits

Figure 7. Strategies for the New York Declaration on Forests Restoration

Note: ANR = assisted naturally regenerated restoration, EP = enrichment planting, PLR = preventing logging reentries and RIL = reducing impact logging

52 of 52

Introduction

Early Active Fire Detection

In recent years, wildfires have become major devastating hazards that affect both public safety and the environment. Thus,agile detection of the wildfires is desirable to suppress the wildfires in the early stage. Owing to the high temporal resolution, GOES-R satellites offer capabilities to rapidly obtainimages every 15 minutes. By utilizing the GOES-R data offered by Google Earth engine, I would like to provide a way to accurately and instantly detect active fires.

Goal

  • Detect the wildfire as early as possible using GOES-R satellite image and compare the detection time with VIIRS 375m active fire product
  • Provide more accurate early active fire mapping and burned area mapping compared to GOES-R active fire product

Method

Due to the reason that GOES-R has 2km spatial resolution, to achieve accurate purpose, the image is resampled to 375m spatial resolution. For each pixel, a window of 11 by 11 is selected to form the time-series and to be fed into the GRU-Network to generate the output. As a compromise to accuracy requirements because of the interference of the smoke and cloud, the max-aggregated image is used in inference.

Steps:

  • Resampling to higher resolution and aggregate images every 3 hour to reduce the interference of cloud and smoke.
  • Generate time-series for each pixel using a window of 11 by 11.
  • Generate the label using VIIRS 375m active fire product
  • Training the deep GRU-network
  • Inference using aggregated imaged every hour to achieve time requirements.

Results

The output of the GRU-Network is overlaid on a Sentinel-2 images and the progression of the active wildfire is clearly visualized. Also the detection of the wildfire is earlier than VIIRS 375m active fire product for our study areas.

The Accuracy for the proposed method is also higher than the baseline GOES-R active fire product.

Conclusion

Conclusions

  • The proposed method provides good performance in early detection of the wildfire compared to the widely used VIIRS active fire products.
  • The proposed method also shows frequent and more accurate mapping of the active fire and burned area mapping than the GOES-R active fire product.
  • The paper is now published in IGARSS 2021 proceedings, for more information, please refer to the conference paper.

References

EARLY DETECTION OF WILDFIRES WITH GOES-R TIME-SERIES AND DEEP GRUNETWORK

Yu Zhao, zhaoyutim@gmail.com