Inventorization of Prospective Areas in Need of Ecological Restoration

​

April 2025

A CoRE Stack project

Indian Institute of Technology Delhi

​

​

Apoorva Dewan, Ojasv Bansal, Raman Kumar, Aaditeshwar Seth

Introduction

• Two methods are generally used to identify areas that need different types of restoration support:
• Maps on land-use, soil erosion, rainfall intensity, population density, etc. are combined together through +ve/-ve rules defined by experts plus local contextual knowledge to identify specific areas.
• Change detection using longitudinal maps available over time are used to identify areas that have degraded during this time and can benefit from restoration support.
• Both types of mapping methods suffer from problems like the maps may be available for only a limited historical period, the fidelity of land-use and other classes may be inadequate to build rules that can capture the diversity of landscapes, and the correctness of the rules to determine areas needing restoration may need rigorous peer validation.
• We are trying to experiment with a new method that takes as input an example of a site already identified for restoration by experts, and tries to find more such sites.
• Our hypothesis is that starting with a specific example will help narrow down the search for more such sites, rather than work top-down to build a logic that covers all possible types of sites.

Solution approach

Define as specific a rule as possible based on various maps to identify this site. Available CoRE stack layers include LULC, Terrain, Fire occurrence, Rainfall, Elevation, Tree canopy density, Change classification

Apply the rule to identify other sites in close proximity. Do post-processing to remove small patches or define additional filters.

Layers available on the CoRE Stack for Step 2 - rule-based classification

• LULC * @ 10m resolution: Shrubs, barren areas, tree cover…
• Cropping intensity * @ 10m resolution: Single/Double/Triple cropping
• Tree canopy density * @ 25m resolution: Above/below median for each ACZ
• Terrain @ 30m resolution: Flat-top hills, deep valleys, plains…
• Drainage lines and stream orders @ 30m resolution
• Catchment area @ 30m resolution: Unit in hectares, of size of catchment
• Other external layers that are integrated
• SRTM DEM: Elevation, Slope @ 30m resolution
• MODIS **: Fire occurrence @ 500m resolution
• GLC *** change detection @ 30m resolution

​

* Annually, 2017 onward. ** Annually, 2003 onward. *** 5-yearly, 1985 onward.

Example: LULC Layer, 10m resolution, annually 2017-

Example map: Angul, Odisha

• Developed using machine learning methods trained on groundtruth local to India, collected from across various ACZs in the country.
• Accuracy between 80-96% for different classes

How we use it in our model

• Apply as a mask for ‘LULC == Shrubs_scrubs’, for example.

Example: Terrain Layer, 30m resolution

Topological Position Index

Compare a pixel’s elevation to the mean elevation of its surroundings, at different distances. Can distinguish many features based on short-distance and long-distance TPI

How we use it in our model

• Apply as a mask for ‘terrain == Mesa tops’, for example.

Example: Fire occurrences, 500m resolution

• Frequency criteria:
• Number of months of fire occurrences between (start_date, end_date)

​

• How we use it in our model
• NASA’s MODIS dataset to extract fire occurrence frequency.
• Apply as a mask for ‘no. of fire_occurences > 1’.

0 fires

10 fires

Between 2010 to 2024

Example: Elevation, 30m resolution

• NASA’s SRTM (Shuttle Radar Topography Mission) dataset to extract elevation data.
• Apply as a mask, for example, ‘500 < elevation < 1000’ to get our required regions.

500m

1000m

Example: Tree canopy density, 25m resolution

• Built using machine learning on optical and SAR satellite data, trained using groundtruth provided by the GEDI mission.
• ACZ-specific models are trained.
• Classes: Above/below median canopy density for each ACZ.
• Outputs are available annually, since 2017

How we use it in our model

• Apply as a mask for ‘canopy density == below median’, as an example

Example: Change detection, 30m resolution

• The GLC_FCS30 data product released in 2024 (paper) provides a global land cover map from 1985 onward.

Degraded areas

Non-degraded areas

How we use it in our model

• Apply as a mask for ‘areas forested in 2000 and deforested in 2024’, as an example

Step 2: Rule-based classification. Examples.

Define as specific a rule as possible based on various maps to identify this site. Available CoRE stack layers include LULC, Terrain, Fire occurrence, Rainfall, Elevation, Tree canopy density, Change classification

Apply the rule to identify other sites in close proximity. Do post-processing to remove small patches or define additional filters.

Rules used to identify degraded forest areas in Angul

(LULC == "Shrub/Scrub") AND (TERRAIN == "Broad Flat Areas" OR "Broad Open Slopes") AND (RAINFALL > 1200mm) AND (150m <ELEVATION <300m) AND (CHANGE DETECTION == “deforested areas”)

Rules used to identify Orans in Rajasthan

(LULC == "Shrub/Scrub" OR "Barren Lands") AND (TERRAIN == "Broad Flat Areas" OR "Upland incised drainages" OR "Broad Open Slopes") AND (RAINFALL < 400mm)

​

Rules used to identify barren hilltops in the Sahyadris

(LULC == "Shrub/Scrub") AND (TERRAIN == "Mesa tops" OR "Local ridges") AND (NUM_FIRE_OCCURRENCES_2010_2024 > 1) AND (ELEVATION > 800m)

Rules used to identify degraded areas in the Sholas

(LULC == "Barren Lands" OR "Single Cropping" OR "Shrub-Scrub") AND (TERRAIN == "Broad Slopes" OR "Upper Slopes" OR "Local Ridges" OR "Mesa Tops") AND (RAINFALL > 1200mm) AND (ELEVATION > 1500m) AND (CHANGE DETECTION == “deforested areas”)

Step 3: Identify other similar sites.

Define as specific a rule as possible based on various maps to identify this site. Available CoRE stack layers include LULC, Terrain, Fire occurrence, Rainfall, Elevation, Tree canopy density, Change classification

Apply the rule to identify other sites in close proximity. Do post-processing to remove small patches or define additional filters.

Step 4: Segmentation.

Define as specific a rule as possible based on various maps to identify this site. Available CoRE stack layers include LULC, Terrain, Fire occurrence, Rainfall, Elevation, Tree canopy density, Change classification

Apply the rule to identify other sites in close proximity. Do post-processing to remove small patches or define additional filters.

Use segmentation algorithms like SNIC to grow the identified sites to larger homogenous regions.

Final outputs for degraded forest areas in Angul

Step-2

Step-3

Step-4

Final outputs for Barren Land areas in Orans

Step-2

Step-3

Step-4

Since large landscapes already match the criteria, further expansion may not be needed.

Final outputs for degraded forest areas in Sholas

Step-2

Step-3

Step-4

Final outputs for barren hilltops in the Sahyadris

Step-2

Step-3

Step-4

Call for participation

We invite organizations involved in restoration projects to share a few details about these projects, including the precise location of the project.

• Give details: https://forms.gle/kUYRwhre61QdXUUQ8

We will try to identify similar areas in need of restoration support and verify these with the organizations. As we improve our methods, we might be able to build a taxonomy of restoration projects and an inventory for each type of project.

This inventorization exercise could help raise funds, build templates for new types of works under MGNREGA, and lead to further work on building a methodology to track the impact of each type of work.

We have deliberately started with a simple methodology so as to learn and understand the area of restoration planning in a better way. We plan to move to advanced methods involving vision transformers in the future.

• To get in touch, write to us at: contact@core-stack.org or aseth@iitd.ac.in