1 of 85

Guidelines to Temporal and Spatial Assessment of Collections

2 of 85

  1. Methods and innovations
  2. Comparisons between collections
  3. Current collection
  4. Comparisons with reference map/data
  5. Improvement challenges and planned innovations for the next collection

Collection assessment

3 of 85

  • Methods and innovations
    1. Overview *
    2. Users or scientific committee comments and demands and MapBiomas’ team response

  • Comparisons between collections:

What are the improvements and limitations of the current collection compared to the two previous collections?

    • Comparison of the mapped area of ​​each class throughout the time series *
    • Spatial comparison of the mapped area of ​​each class by tiles or regions
    • Mapped class agreement *
    • Mapped area comparison using Sankey Diagram *
    • Global accuracy comparison*
    • Omission and commission errors*
    • Comparison of deforestation and secondary vegetation

Collection assessment

* Priority analysis

4 of 85

  • Methods and innovations
  • Comparisons between collections
  • Current collection:

What do the spatiotemporal patterns of mapped area and accuracy analysis throughout data processing tell us about the methodological steps and data quality?

    • Impact of spatial and temporal filters and themes integration *
    • Spatiotemporal variation of accuracy
    • Temporal analysis: number of changes, number of classes and trajectories for each mapped class*
    • Deforestation and secondary vegetation recovery by class and age

Collection assessment

* Priority analysis

5 of 85

  • Methods and innovations
  • Comparisons between collections
  • Current collection
  • Comparisons with reference map/data (eg. TERRACLASS, FBDS, GLAD, Inventário, IBGE, PRODES): Considering differences in class definitions and spatial resolution, to what extent are the results in accordance with other data sources?
    • Total agreement and disagreement*
    • Understanding the disagreement
    • Identifying regions of over and underestimation*
    • Quantitative comparisons of mapped area*
    • Accuracy analysis
    • Comparisons with proxy data
    • Comparisons of deforestation data
  • Improvement challenges and planned innovations for the next collection*

Collection assessment

* Priority analysis

6 of 85

  • Always include a sentence that summarizes the main insight provided by that figure.
  • Don't forget to include axis titles AND measurement units in charts in READABLE font size.
  • Don't forget to include self-explanatory legends in the figures. The reader should be able to understand the figure without any explanation or information provided in previous slides.
  • When comparing collections, only take into account the current one and the last two.

General Recommendations

7 of 85

  1. Methods and innovations

8 of 85

  • Samples filtering
    • SAD Cerrado (2020-2022) and MapBiomas Alerta (2019-2022) masks

  • Complete regionalization of the classification flow
    • Sample selection and screening
    • Hyperparameters
    • Feature Space
    • Classification

  • Reassessment of critical post-processing steps
    • Incidence filter
    • Temporal filter

  • Expansion and refinement of the Rock Outcrop class
    • Expanded territorial scope
    • Complementary samples

INNOVATIONS, GAINS AND COMMITMENTS

Savanna - COLLECTION 8.0

Methodological workflow

1a. Overview*

Savanna Team

Present a flowchart of the methods and highlight the innovations in relation to the previous collection

9 of 85

1b. Users’ comments and demands and MapBiomas’ team response

Atlantic Forest Team

Inclusion of the mapping of MG rock fields following a request by email.

Present examples of improvements motivated by users comments and demands

10 of 85

2. Comparisons between collections

What are the improvements and limitations of the current collection compared to the two previous collections

11 of 85

C6.0: 83

C5.0: 83.8

C4.1: 82.9

C3.1: 81.4

Média

C6.0: 74.9

C5.0: 74.8

C4.1: 72.6

C3.1: 71.5

  • Level-1

Reduced accuracy in C6 compared to C5, changes of classes and legend levels (eg. Silviculture)

  • Level-2

Not comparable

(igual L3 na C6)

  • Level-3

Increase in accuracy throughout the collections, but with lower accuracy at the beginning of the series

2a. Global accuracy comparison*

Atlantic Forest Team

Compare the global accuracy throughout the time series between collections in different legend levels

12 of 85

Collection 8 - Level 1

Better global accuracy than other col. : 84.7%

Allocation errors: 14.5% (11.8% Col. 7.1)

Quantity errors 0.8% (3.9% Col. 7.1)

Collection 8 - Level 3

Lower global accuracy than col. 7.1: 76.1%

Allocation errors: 13.3% (15.6% Col. 7.1)

Quantity errors 10.6% (8.1% Col. 7.1)

Savanna Team

2a. Global accuracy comparison*

Compare the global accuracy throughout the time series between collections in different legend levels

13 of 85

Highlights of the 7.1 collection

Global Accuracy

Level 3

col 7.1 < col 5 e 6

col 7.1 > col 7

All collections above 85% accuracy

Accuracy at the beginning of the series is lower, but with a variation between 82% and 87% considering ALL classes at level 3

Atlantic Forest Team

2a. Global accuracy comparison*

Compare the global accuracy throughout the time series between collections in different legend levels

14 of 85

2a. Global accuracy comparison

Compare the global accuracy of regions or tiles throughout the time series between collections**

Savanna Team

** Script available in analysis 3C; Regions may be substituted by tiles (“cartas/ super cartas”) available at projects/mapbiomas-workspace/AUXILIAR/cim-world-1-250000

15 of 85

Increase of the anthropic areas in Col. 8

Mosaic +4,2 Mha

Pasture +5,4 Mha

Col 7.1

Col 8.0

Savanna mapped as “mosaic of uses” and pasture

2b. Mapped area comparison using Sankey Diagram*

Savanna Team

Analysis of changes in pixels’ class assignment between collections: proportion of pixels that changed classes and new classes

16 of 85

https://sites.google.com/alumni.usp.br/mapbiomas-tools/sankey_71e8?authuser=0

Considering only pixels that changed class

Same as previous analysis, but excluding stable pixels in order to better visualize the proportion of change to each new class in the current collection

2b. Mapped area comparison using Sankey Diagram*

17 of 85

Floriano (PI)

Blue = Collection 8,

Red = Collection 7.1

Green = Both Collections

Mosaic of Uses

Mosaic of uses areas that became pasture in Collection 8

Savanna Team

2c. Mapped class agreement (between current and previous collections)

COLLECTION 7.1 - COLLECTION 8.0 (2021)

Present a map showing which pixels are mapped in a specific class in both current and previous collection or in each of them

18 of 85

Agreements por Classe 2018

Savanna Formation

2c. Mapped class agreement (between current and previous collections)

Quantify the area of specific classes in both current and previous collection or in each of them throughout the time series

Caatinga Team

19 of 85

INNOVATIONS, GAINS AND COMMITMENTS

Savanna - COLLECTION 8.0

Collection 7.1

Region 30 (2021)

2c. Mapped class agreement (between current and previous collections)

Savanna Team

Collection 8

Present examples of improvements and challenges in mapping specific classes

Decrease in commission errors of grasslands and savanna formations in non vegetated areas

Decrease in commission errors of native vegetation in relief shades

20 of 85

Silviculture

Collection 6 assessment

Collection 5

Collection 6

2c. Mapped class agreement (between current and previous collections)

Agriculture Team

Present examples of improvements and challenges in mapping specific classes

21 of 85

Agriculture and forestry map overlay between three collections, grouping temporary crops, perennial crops and forestry.

Comparison of C6, C7.1 and C8

2020

C6

C7

C8

Agreements por Classe 2018

2c. Mapped class agreement (between current collection and the previous two)

Present a map overlay of the current collection and the previous two for a specific class

Agriculture Team

22 of 85

Agriculture and forestry map overlay between three collections, grouping temporary crops, perennial crops and forestry (“level 3”).

Comparison of C6, C7.1 and C8

2020

1985

C6

C7

C8

Agriculture Team

Agreements por Classe 2018

2c. Mapped class agreement (between current collection and the previous two)

Based on the previous map, quantify the proportion of common area between the three collections by regions/biomes

Overlapping area between collections - sum of the entire time series

23 of 85

STABILITY BETWEEN COLLECTIONS

COLLECTIONS 6.0, 7.1 and 8.0 (2019)

Collection 6

Collection 7.1

Collection 8

Number of classes

Forest

Forest

Forest

1 - Concordant

5 –

Very discordant

Forest

Grassland

Agriculture

Grassland

Forest

Forest

Grassland

Grassland

Forest

2 – Recent concordance

3 – Recent discordance

Forest

Grassland

Forest

4 - Discordant

Legenda

HIgher uncertainty classification

1

2

2

2

3

2c. Mapped class agreement (between current collection and the previous two)

Savanna Team

Considering the current and the previous two collections, map and quantify the occurrence of pixels’ classifications disagreement between collections using agreement categories

24 of 85

Pampa Team

Agreements por Classe 2018

2c. Mapped class agreement (between current collection and the previous two)

Based on the previous map, quantify the proportion of common area between the three collections

Collections 7.1, 8 and 9

91,6%

Col 7.1

Col 8

Col 9

Venn Diagram

0,2%

3%

4,1%

1,1%

91,6%

Mean values:

17,8 Mha

Concordant

4,1%

0,8 Mha

Recent Concordant

0,2%

0,05 Mha

Very discordant

1,1%

0,2 Mha

Discordant

3%

0,6 Mha

Recent Discordance

25 of 85

2c. Mapped class agreement (between current collection and the previous two)

Considering the current and the previous two collections, map and quantify the occurrence of pixels’ classifications disagreement between collections using agreement categories

Col 7

Col 6

Col 5

Pampa Team

Con

Concordant

Recent concordance

Recent discordance

Discordant

Very discordant

26 of 85

STABILITY BETWEEN COLLECTIONS

COLLECTIONS 6.0, 7.1 and 8.0 (2019)

Nova Colinas (MA)

Considering 2019 for all collections

Savanna Team

2c. Mapped class agreement (between current collection and the previous two)

Considering the current and the previous two collections, map and quantify the occurrence of pixels’ classifications disagreement between collections using agreement categories

27 of 85

Between the last 3 collections, 16,8% agree on two or no collections, of which 0,8% do not agree in any pair of collections (class 5)

It is important to improve the class 3 - 6,7% ( recent discordance)

2c. Mapped class agreement (between current collection and the previous two)

Savanna Team

STABILITY BETWEEN COLLECTIONS

COLLECTIONS 6.0, 7.1 and 8.0 (2019)

Year 2019

Considering the current and the previous two collections, map and quantify the occurrence of pixels’ classifications disagreement between collections using agreement categories

28 of 85

4,7%

23,8%

2c. Mapped class agreement (between current collection and the previous two)

Pampa Team

Agriculture

Wetland

Same as previous analysis, but quantifying the area in each agreement/disagreement category throughout the time series by class

29 of 85

Diferença de Área Col 7.1 vs Col 8.0�Vegetação Nativa (3, 4, 11 e 12)

MATOPIBA: Maiores mudanças de áreas naturais em relação à Col 7.1

2d. Spatial comparison of the mapped area of ​​each class by tiles or regions

Savanna Team

Area difference Col 7.1 vs Col 8

Native vegetation (3,4,11 and 12)

MATOPIBA: greaters changes in natural areas compared to Col. 7.1

Provide a quantification by tiles or regions of increase or decrease of each class’ area between collections

30 of 85

2d. Spatial comparison of the mapped area of ​​each class by tiles or regions

Urban Areas Team

Provide a quantification by charts of agreement or disagreement of each class’ area between collections

1985

2020

disagreement

agreement

% of the area

% of the area

31 of 85

2d. Spatial comparison of the mapped area of ​​each class by tiles or regions (by states)

COMPARISON BETWEEN MAPPED AREAS IN COLLECTIONS 7 & 8

More areas in Col. 7

More areas in Col. 8

Stability

MT -59.65204, -14.84975

Only Col.8

Only Col.7

Both collections

0

500km

- Commission error adjustment in Mato Grosso areas

-24.01

29.52

11.48

0.8

1.29

0.23

0.12

0.05

6.88

0.14

-0.9

0.3

9e-3

92e-2

0.32

1.19

89e-2

0.18

71e-2

0.67

34e-2

-2e-7

0

0

0

0

0

Area in Mha

Mining Team

Compare the mapped area between collections by states or regions

32 of 85

Forest Formations

Savanna Formations

Grasslands

Area(ha)

2021�Collection 7.1: 28.052.538 ha

Collection 8.0: 24.826.613 ha

2021Collection 7.1: 60.513.568 ha

Collection 8.0: 56.558.461 ha

2021Collection 7.1: 10.496.488 ha

Collection 8.0: 7.636.620 ha

-3.225.925 ha

-3.955.107 ha

-2.859.868 ha

COMPARISON OF MAPPED AREAS (Col 7.1 and Col 8.0)

2e. Comparison of the mapped area of ​​each class throughout the time series

Savanna Team

Compare the total mapped area of each class throughout the time series between collections

33 of 85

4,3 Mha

3,2 Mha

2,6 Mha

Pampa Team

2e. Comparison of the mapped area of ​​each class throughout the time series (emphasizing the difference between classes)

Compare the total mapped area of each class throughout the time series between collections

34 of 85

2e. Comparison of the mapped area of ​​each class throughout the time series (by region)

?

A

A#

R3 - Campanha (oeste)

R7 - Zona Costeira

R5 - Fronteira Oeste

R2 - Campanha (leste)

Regions of Pampa BR

R1 -Serra do Sudeste

Pampa Team

Compare the total mapped area of each class throughout the time series between collections

35 of 85

3 – Forest Formation

Pampa Team

2e. Comparison of the mapped area of ​​each class throughout the time series (by region)

Area difference Col 7.1-Col 6

Compare the total mapped area of each class throughout the time series between collections

36 of 85

Urban area comparison between Col. 5, Col. 7.1, Col.8 and IBGE (2019)

Tiles not mapped in MapBiomas collections that have small urbanized area.

Attention: tiles with worsening coincidence with IBGE throughout the collections and more than 20,000 ha of Urbanized Area (following IBGE).

Compare the class mapped area between collections considering a reference data

Urban Areas Team

2e. Comparison of the mapped area of ​​each class throughout the time series (by tiles, using reference data)

37 of 85

2f. Omission and commission errors*

Pampa Team

Errors - Grasslands

Present charts or maps that compare omission and commission errors for each class between collections throughout the time series

38 of 85

Col 7.1

Col 8

Errors by class: how much, with what class, when?

139

Mean number

of points

55

25

1200

677

29*

49

Pampa Team

2f. Omission and commission errors*

Present charts or maps that compare omission and commission errors for each class between collections throughout the time series

39 of 85

Desmatamento

  • 30% reduction in the area of ​​secondary vegetation between col 8 and col7.1, mainly the result of the minimum area filter that reduces “false regeneration”
  • Change in regeneration calculation
  • 7.1 - change of at least 1ha from one year to the next
  • 8 - minimum 1ha change in the sum of all years in the series

Avg. 321mil ha

Avg. 447mil ha

  • Reduction in noise from secondary vegetation reflects a decrease in deforestation values
  • We are closer to the value published by SOS Mata Atlântica, which is +-50 thousand ha / year

Atlantic Forest Team

2g. Comparison of deforestation and secondary vegetation

Compare the suppression of primary and secondary vegetation throughout the time series between collections

40 of 85

In 2021, 26% of natural vegetation is Secondary Vegetation in collection 7.1

26%

74%

23%

77%

COLEÇÃO 7.1

COLEÇÃO 8

In 2021, 23% of natural vegetation is Secondary Vegetation in collection 8

Redução da contribuição da vegetação secundária na área total de floresta da MA

Atlantic Forest Team

2g. Comparison of deforestation and secondary vegetation

Compare secondary vegetation area throughout the time series between collections

41 of 85

3. Current collection

What do the spatiotemporal patterns of mapped area and accuracy analysis throughout data processing tell us about the methodological steps and data quality?

42 of 85

https://public.flourish.studio/visualisation/15089325/

3a. Impact of spatial and temporal filters and themes integration*

Indonesia Team

Quantify the impact of filters on the mapped area of ​​each class throughout the historical series

43 of 85

3a. Impact of spatial and temporal filters and themes integration*

Coastal Zone Team

Remote Sens. 2019, 11, 808; doi:10.3390/rs11070808

Show examples of the effects of filters application

44 of 85

Map which pixels were originated by filters, the integration phase or original classification

Buriti Bravo (PI)

Incidence filter (green) influence in the final classification

Mambaí (GO)

3a. Impact of spatial and temporal filters and themes integration*

Savanna Team

COLLECTION 8.0

45 of 85

Of the entire Savanna classification, 9.7% of the pixels come from post-classification filters and 3.4% come from integration

PIXEL’s ORIGIN ANALYSIS

COLLECTION 8.0

Year: 2022

Savanna Team

3a. Impact of spatial and temporal filters and themes integration*

Quantify the proportion of pixels originated by filters, the integration phase or original classification

46 of 85

Pre-filters

Post-filters

Integrated

NUMBER OF CLASSES PER PIXEL

NUMBER OF CLASSES PER PIXEL

NUMBER OF CLASSES PER PIXEL

In the pre-filters stage, the classification has up to 6 classes per pixel. Post-processing filters have a significant effect on the number of classes per pixel.

After integration, 50% of pixels have only 1 class

Show the effect of filters application on the number of classes per pixel throughout the time series

3a. Impact of spatial and temporal filters and themes integration

Savanna Team

47 of 85

2022

Classification

Filtered Classification

Show the effects of applying filters on classes using a Sankey Diagram

3a. Impact of spatial and temporal filters and themes integration

Pampa Team

48 of 85

Compare the effects of filters application between regions and throughout the time series

3a. Impact of spatial and temporal filters and themes integration

Savanna Team

Cartogram of the most active filter, by region

sum of the series 1985-2022

https://github.com/mapbiomas-brazil/cerrado/tree/master/3-integration/pixelSource

Percentage change on the map, per year

*Considera .remap para as classes [3, 4, 11, 12, 21, 25, 33]

4%

4.6%

0.1%

4.5%

1.9%

3.7%

5.9%

0.1%

3.3%

1.2%

2.9%

3.3%

5%

0.1%

3.6%

5.3%

3.6%

5.4%

0.1%

3.2%

1.3%

3.2%

17% (∓1%/ano) do mapa alterado pelos filtros

3,2% (∓0,8%/ano) do mapa alterado pela integração (Nat -> Ant)

3a. Impact of spatial and temporal filters

The temporal filter is the most active in 68% of the regions

16% incidence (areas with dynamics of use/degradation of native vegetation without conversion - e.g. pasture)

16% integration (areas of older/consolidated use and transition with Atlantic Forest/Amazon)

5% spatial (areas with high heterogeneity and fragmentation of native vegetation)

49 of 85

IMPACT OF POST-CLASSIFICATION FILTERS COLLECTION 9

(1985 - 2023)

MEAN FILTER EFFECT BY CLASS

Grasslands and Agriculture are the classes most changed by filters

Área (Mha)

1,5%

1%

1%

1%

0,6%

0,4%

0,5%

3,1%

1,6%

2%

0,7%

0,3%

2,2%

Compare the effects of filters application considering the classes

3a. Impact of spatial and temporal filters and themes integration

Pampa Team

50 of 85

Compare the effects of filters application between classes and throughout the time series

3a. Impact of spatial and temporal filters and themes integration

Savanna Team

28% (∓10%/year) of water added by GT Água

Spatial Filter 4% (∓0.3%/year) Acts on fragment edges

8% (∓1.4%/ano) added by incidence filter

5% (∓1.6%/ano) added by transversal themes

6.6% (∓1.2%/ano) added by other filters

Frequency filter is the most active

7% (∓1%/year) - It avoids seasonal fluctuations and maintains the structural concept of the class

Temporal filter is the most active 9% (∓1%/ano)

It prevents false transitions to grassland and forest

Temporal filter is the most active 7% (∓1%/year) - It prevents false transitions to savanna

29% (∓3%/ano) adicionado pelos transversais

4% (∓0.9%/ano) of native vegetation added by post-integration filters

Water

Forest

Anthropogenic Use

Wetland

Savanna Formation

Grassland

Non vegetated

51 of 85

Present maps of global accuracy by regions or tiles

3b. Spatiotemporal variation of accuracy

Collection 5 – Global Accuracy

Global accuracy by legend level

Level 1

Level 2

Level 3

Global accuracy

Savanna Team

52 of 85

3b. Spatiotemporal variation of accuracy

Savanna Team

Present maps of global accuracy by regions or tiles throughout the time series

Median 81.9%

Max. median 83.7% - 2012

Min. median 80.1% - 1993

  • Temporal stability of global accuracy
  • Temporal stability of standard deviation
  • Stable spatially explicit patterns

Global accuracy per year - legend level 3

53 of 85

Forest Formation

2020

2.839 points of forest mapped correctly as forest

Omission

  • 211 points of forest mapped as mosaic of uses
  • 65 points of forest mapped as savanna formation

Comission

  • 133 points of pasture mapped as forest
  • 101 points of silviculture mapped as forest

  • Agreement in 82% of the points
  • 10% of omission, mainly in mosaic of uses areas
  • 8% of comission

Comission

Agreement

Omission

3b. Spatiotemporal variation of accuracy

Present maps of omission and commission errors for the main classes

Atlantic Forest Team

54 of 85

OMISSION ERRORS DISTRIBUTION MAPS

AMAZÔNIA

Savanna Formation

Densidades de pontos

Baixa

Alta

3b. Spatiotemporal variation of accuracy

Present heatmaps of omission and commission errors throughout the time series

Amazonia Team

55 of 85

Pixel size: 1 km

Kernel size: 100 km

H3

N of points

Omission

Comission

3b. Spatiotemporal variation of accuracy

Present heatmaps of omission and commission errors throughout the time series

Pasture Team

56 of 85

Atlantic Forest Team

3c. Temporal analysis: number of changes, number of classes and trajectories for each mapped class*

Present gif animations showing class changes throughout the time series

57 of 85

3c. Temporal analysis: number of changes, number of classes and trajectories for each mapped class*

Pasture Team

Pasture Mapping -

Collection 8 (1985 - 2022)

Present gif animations showing class changes throughout the time series

58 of 85

Areas of forest overestimated in the initial years of the series

Pantanal Team

3c. Temporal analysis: number of changes, number of classes and trajectories for each mapped class*

Based on maps of pixels trajectories categories, identify possible inconsistent spatiotemporal patterns

59 of 85

Grasslands

Trajectories assessment

Region 5 - increase the limit to avoid sudden changes in the classification

Region 24 - reduce the limit to avoid noise at the border of grasslands

Noise in the early years of the classification

3c. Temporal analysis: number of changes, number of classes and trajectories for each mapped class*

Atlantic Forest Team

Based on maps of pixels trajectories categories, identify possible inconsistent spatiotemporal patterns

60 of 85

Grasslands

Trajectories assessment

Increase the duration of the period of the moving window post-classification temporal filter in order to remove 2, 3 or 4 years of mosaic of uses or pasture between natural classes

Atlantic Forest Team

3c. Temporal analysis: number of changes, number of classes and trajectories for each mapped class*

Based on pixels trajectories examples, identify possible inconsistent temporal patterns and solutions

61 of 85

3c. Temporal analysis: number of changes, number of classes and trajectories for each mapped class*

Based on maps of pixels trajectories categories, identify possible inconsistent spatiotemporal patterns

Trajectory Toolkit

62 of 85

Forest formation

Savanna Formation

Grassland

Wetland

Perda média de primária

137 mil ha/ano

Perda média de primária

500 mil ha/ano

Perda média de primária

62 mil ha/ano

Perda média de primária

11 mil ha/ano

DEFORESTED AREA BY CLASS (1986 a 2021)

Área (ha)

Área (ha)

LOSS OF NATIVE VEGETATION IN THE SAVANNA

Diferença entre os dados chega a 283 mil ha (1997)

Área (ha)

3d. Deforestation and secondary vegetation by class and age

Compare the filtered (working group) and raw deforestation data, and between classes

Savanna Team

63 of 85

How old is the secondary vegetation when deforestation occurs?

SECONDARY VEGETATION LOSS IN THE SAVANNA (DEFORESTATION WORK GROUP)

In 2020, 31.5% (42 thousand ha) of second. veget. deforestation occurred in vegetation aged 0 to 5 years, an age that represents 68% of the second. veget.

Secondary vegetation over 20 years old accounted for 15% of deforestation

Área (ha)

MATOPIBA

100%

25%

30%

18%

28%

31%

31%

14%

8%

25%

15%

Área desmatada de Veg. secundária

31%

69%

30%

30%

15%

10%

17%

Foto: Rafael Coelho /IPAM

Savanna Team

3d. Deforestation and secondary vegetation by class and age

64 of 85

SECONDARY VEGETATION RECOVERY IN THE SAVANNA

Agriculture → Second. Veg.

Silviculture → Second. Veg.

Mosaic of uses → Secon. Veg.

Pasture → Sec. Veg.

In 2020, the Cerrado had 487 thousand hectares of recovery for secondary vegetation

Of this area, 70% (345 thousand ha) represents regrowth in pasture areas

Área (%)

Ganho médio de secundária: 500 mil ha/ano

Mean

16K ha/year

Mean

12K ha/year

Mean

248K ha/year

Mean

274 k ha/ano

Savanna Team

3d. Deforestation and secondary vegetation by class and age

65 of 85

4. Comparisons with reference map/data

Considering differences in classes definitions and spatial resolution, to what extent are the results in accordance with other data sources?

66 of 85

Terraclass (2013) classes: Natural, Natural non vegetated, Silviculture, Pasture, Agriculture, Mining, Urbanizaded Area and others

MapBiomas Col. 8 (2013)

Terraclass 2013

Concordância/Discordância

55,2%

44%

50,4%

48,9%

68,4%

31,6%

Natural Florestal: 418.789 km²

Natural ñ Florestal: 692.301 km²

Natural Florestal: 249.130 km²

Natural ñ Florestal: 751.177 km²

Veg. nativa

Antrópico

Veg. nativa

Antrópico

Savanna Team

4a. Total agreement and disagreement

Present a map of agreement/disagreement of current collection compared to a reference map for a given year

67 of 85

LabGeo/UFRGS 2015

Col. 8 2015

Discordância

4b. Understanding the disagreement

When a wall to wall reference map is available, identify the confusion between classes using a Sankey diagram

Pampa Team

68 of 85

Perennial crop not mapped by Glad

Large adherence between Glad and MapBiomas maps

Large adherence between Glad and MapBiomas maps

Omission of sugarcane by mapbiomas

A

B

D

A

B

C

D

4c. Identifying regions of over and underestimation*

Agriculture Team

Present a map showing which pixels are mapped in a certain class in both current collection and in a reference map or in each of them

Only GLAD mapped

Only MapBiomas C6 mapped

Glad and MapBiomas mapped

C

Agriculture

Analysis of Collection 6 compared with

69 of 85

COMPARISON WITH REFERENCE DATA

TERRACLASS Savanna - COLLECTION 8.0 (2020)

Mapping Differences

Terraclass vs Col 8.0�Natural Classes (3, 4, 11, 12 e 29)

Vegetation Classes (1, 2)

Difference of 113,572 km² (11%) in the native vegetation class (MapBiomas - Terraclass)

Largest area differences (MapBiomas - Terraclass)

  • Pasture: +103.155 km²
  • Agriculture: -34.787 km²

+Red: Terraclass maps more native vegetation

Terraclass 2020

Collection 8 - 2020

4d. Quantitative comparisons of mapped area (by tiles or regions)*

Savanna Team

If the spatial resolution differs among current and reference classifications, you may present a map of class area difference between them by tiles or regions. Present total numbers and proportions in charts.

70 of 85

ESRI - GLC - ESA- GLAD - COlLECTION 9 (2020)

Legend

  • Water and Agriculture were the classes with the highest agreement.

(MapBiomas x Others)

  • On average, 63% of areas mapped as herbaceous and shrubby vegetation

were also mapped

in other initiatives

(ESRI, GLC, ESA, GLAD)

Alta discordância entre as coleções

Percentage of extra pixels mapped by MapBiomas (class color) or not mapped (dark grey bar)

**O percentual considera todo o universo de pixels mapeados nas duas coleções comparadas. Assim, o percentual de acerto é: 100% - mapeou a mais - mapeou a menos

Mapeia mais ANV

66,7%

63,4%

71,1%

33,2%

92,5%

23,8%

only MB

mapeou a mais

mapeou a menos

only others

4d. Quantitative comparisons of mapped area*

Present the proportion of pixels mapped by MapBiomas only and by reference map only by class in charts.

ESRI - GLC - ESA- GLAD - COLLECTION 9 (2020)

Pampa Team

71 of 85

4d. Quantitative comparisons of mapped area (throughout the time series)*

Agriculture Team

Compare mapped area estimates between current classification and reference data throughout the time series

72 of 85

4e. Accuracy analysis

Coastal Zone Team

Remote Sens. 2019, 11, 808; doi:10.3390/rs11070808

Present accuracy estimates of current collection based on reference maps.

73 of 85

4e. Accuracy analysis (by states)

Agriculture Team

Present accuracy estimates of current collection based on reference maps.

74 of 85

TERRACLASS - Savanna - Temporary crops

Agriculture Team

4e. Accuracy analysis

Compare accuracy estimates between current collection and reference maps using the same set of test points

75 of 85

4f. Comparisons with proxy data

Urban areas Team

Compare mapped area with other variables data may provide clues of whether spatial and temporal patterns are consistent

76 of 85

LOSS OF NATIVE VEGETATION IN THE SAVANNA

(REFERENCE DATA)

Área (km²)

Dado

Abrangência temporal

Calendário

Média de Desmatamento Anual

GT Desmat.

1987 a 2021

Civil

8,2 mil km²

PRODES

2001 a 2023

Civil

13,5 mil km²

DETER

2017 a 2024

Ano Referência

5,5 mil km²

MB Alerta

2019 a 2023*

Civil

6,5 mil km²

SAD Cerrado

2022 a 2023

Civil

9,3 mil km²

IBAMA

2008 a 2010

Ano Referência

7,1 mil km²

Maior diferença

(GT Desmatamento e PRODES)

21 mil Km²

Os produtos indicam aumento do desmatamento nos últimos anos

*Coleção 8 e soma primário e secundário

* dados até 01/04/24

Anual mean (2001 à 2022)

GT: 6,4 mil km²�Prodes: 13,5 mil km²

4g. Comparisons of deforestation data

Compare deforestation data throughout the time series with other data sources, taking into account differences of deforestation definition, spatial resolution and analized images’ date

Savanna Team

77 of 85

5. Improvement challenges and planned innovations for the next collection

78 of 85

INNOVATIONS, GAINS AND COMMITMENTS

SAVANNA - COLLECTION 8.0

Region 33 (2021)

Collection 7.1

Collection 8

For improvement:

“Loss” of small forest fragments mapped in the collection 7.1 but not in the collection 8

Savanna Team

5. Improvement challenges and planned innovations for the next collection*

Present maps, charts, flowcharts and bullet points to show major challenges, planned innovations and new methodological tests

79 of 85

CONSIDERATIONS / NEXT STEPS

  • Assessment of deforestation data and secondary vegetation dynamics
    • Does the data make sense for the Wetland class?

  • Check secondary vegetation recovery information
    • Check secondary vegetation recovery information

  • Investigation of the time series and transitions
    • Do the outliers make sense?
    • What improvements can be applied at the end of the series?

  • Collection 9 improvements [in progress]
    • Improved sample filtering and classification flow

Savanna Team

5. Improvement challenges and planned innovations for the next collection*

Present maps, charts, flowcharts and bullet points to show major challenges, planned innovations and new methodological tests

80 of 85

Few stable areas

Main challenges for Collection 8

  • Detailed flood pulse information;
  • Forest classification improvements;
  • Reduce confusion between natural grassland vegetation and exotic pastures.

Pantanal Team

5. Improvement challenges and planned innovations for the next collection*

Present maps, charts, flowcharts and bullet points to show major challenges, planned innovations and new methodological tests

81 of 85

New regionalization proposal based on depositional systems and sub-basin

adjustments in sample balance for each region

Planned methodological innovations

Anual classification

Pantanal Team

5. Improvement challenges and planned innovations for the next collection*

Present maps, charts, flowcharts and bullet points to show major challenges, planned innovations and new methodological tests

82 of 85

Challenges and next steps

  1. Agriculture:
    • Reduce omissions in coffee mapping.

  • Silviculture:
    • Obtain reference maps with partners (Forestry Dialogue as intermediary);
    • Reduce omission;
    • Map new areas.

  • Irrigation:
    • Apply post-processing filters to individual pivots;
    • Maintain id of each pivot over time;
    • Calculate Evapotranspiration for all pivots and store this data;
    • Obtain start and end dates for harvests according to the timesat definition;
    • Reduce the number of uncalculated pivots.

Agriculture Team

5. Improvement challenges and planned innovations for the next collection*

Present maps, charts, flowcharts and bullet points to show major challenges, planned innovations and new methodological tests

83 of 85

Reference Map vs Raw MapBiomas C6 results

Reference Map vs New Sampling Methodology Results

Cena 218-75

Perennial Crop

Coffee

MapBiomas C6

MapBiomas C6

New Methodology

New Methodology

Agriculture Team

5. Improvement challenges and planned innovations for the next collection*

84 of 85

Innovations Collection 8 - Flooded Areas (in testing phase)

Amazon Team

5. Improvement challenges and planned innovations for the next collection*

85 of 85

Available Scripts, Tools and Guides