cs.CVFeb 19, 2026
SaveTree crop mapping of South America reveals links to deforestation and conservation
Organizations: Google DeepMind, Zürich, Switzerland · EcoVision Lab, DM3L, University of Zurich, Zürich, Switzerland · University of Copenhagen, Copenhagen, Denmark · Google Research, Seattle, WA, USA ·
Abstract
Monitoring tree crop expansion is vital for zero-deforestation policies like the European Union's Regulation on Deforestation-free Products (EUDR). However, these efforts are hindered by a lack of highresolution data distinguishing diverse agricultural systems from forests. Here, we present the first 10m-resolution tree crop map for South America, generated using a multi-modal, spatio-temporal deep learning model trained on Sentinel-1 and Sentinel-2 satellite imagery time series. The map identifies approximately 11 million hectares of tree crops, 23% of which is linked to 2000-2020 forest cover loss. Critically, our analysis reveals that existing regulatory maps supporting the EUDR often classify established agriculture, particularly smallholder agroforestry, as "forest". This discrepancy risks false deforestation alerts and unfair penalties for small-scale farmers. Our work mitigates this risk by providing a high-resolution baseline, supporting conservation policies that are effective, inclusive, and equitable.
Figures & tables
Figure 1: Tree crop map for South America. The main map displays a hexagon-based overview (80km per hexagon side), where brighter colors indicate higher tree crop density. Detailed maps a–d show zoomed-in views of the tree crop probability in the tree crop regions of (a) Argentina (mainly apple orchards), (b) Brazil (large-scale coffee plantations), (c) Peru (palm plantations), and (d) Colombia (smallholder coffee farms). (e) Tree crop area per country.
Figure 2: Tree crops in the forest cover loss areas. a) Tree crop area within forest cover loss zones (left axis, ha) and its proportion relative to total tree crop area (right axis, %) by country. While higher tree crop area generally corresponds to greater overlap with forest cover loss, Peru stands out as an outlier with a high relative proportion despite a smaller absolute area. b) Temporal patterns of tree crops within forest cover loss in South America. The line chart shows the annual tree crop area inside forest cover loss per country. c) Example of large-scale oil palm expansion in Ucayali, Peru.
Figure 3: Tree crops in the Protected Areas. a) Decrease in tree crop area with increasing distance from the boundary into Protected Areas. b) Increase in tree crop area with increasing distance outward from the boundary of the Protected Areas. c) Examples of selected Protected Areas and PA category names. The red rectangles are Protected Areas, white indicates the predicted tree crop areas while black indicates non-tree crop areas.
Figure 4: Addressing EUDR in comparison of JRC maps. a) Comparison of tree crop representation in version 1 and version 2 of the JRC Global Forest Cover 2020 map. b) The bar chart shows the distribution of JRC’s Global Forest Type 2020 classes within our mapped tree crop areas. c) Zoomed-in examples from Colombia comparing our tree crop map with the JRC Global Forest Cover 2020 (v2). Orange indicates areas classified as non-forest by JRC, while white indicates those classified as forest. Large oil palm plantations are generally mapped as non-forest (windows 1–3 with latitude and longitude: [10.672, –74.240], [7.264, –73.889], [3.916, –73.385]), whereas smaller fields, such as coffee, are often classified as forest (windows 4–5 with latitude and longitude [5.321, –75.733], [2.499, –76.042]).
| Reference | Str. | |||||
|---|---|---|---|---|---|---|
| Stratum | Tree Crops | Non-Tree Crops | Total | Str. area [ha] | Str. weight | |
| M a p | Tree Crops | 1277 | 272 | 1549 | 8,736,814 | 0.0047 |
| Non-Tree Crops | 4 | 1728 | 1732 | 1,816,066,434 | 0.9801 | |
| Buffer | 46 | 534 | 580 | 28,110,734 | 0.0152 | |
| Total | 1327 | 2534 | 3861 | 1,852,913,982 | 1.000 | |
Table 1: Error matrix for tree crop map assessment. Rows represent map predictions; columns represent reference labels. is the stratum weight based on the mapped areas.
| Country | UA | PA | Initial area | Adjusted area | (%) |
|---|---|---|---|---|---|
| Brazil | 90 | 62 | 4,609,541 | 6,644,558 | 0.528 |
| Colombia | 70 | 79 | 1,557,885 | 1,390,107 | 1.358 |
| Ecuador | 73 | 67 | 834,422 | 916,727 | 3.355 |
| Chile | 76 | 68 | 871,309 | 967,329 | 0.922 |
| All countries ( ) | 83 | 66 | 7,873,157 | 9,908,008 | 0.711 |
Table E1 : Tree crop area assessment by country. UA = User’s Accuracy and PA = Producer’s Accuracy (after adjustment). The initial area is derived directly from the map, while the adjusted area reflects post-adjustment estimates. denotes the stratum weight of tree crops as a percentage of total land area.
Figure A1 : Hexagon map of tree crop areas inside forest cover loss per country. Each hexagon covers area and darker hexagon indicate more tree crop areas inside forest cover loss. The line chart shows the annual tree crop area inside forest cover loss per country.
Figure A2 : Examples of visually interpreted tree crop areas within forest cover loss. (a) and (b) show examples of likely forest-to-tree-crop conversion in Ecuador and Colombia, respectively, while (c) illustrates tree crop replanting in Brazil. Images are historical high-resolution satellite images from Google Earth.
| Name | Description |
|---|---|
| SDPT Tree Crops | A collection of tree crop sources, including both, manual annotations and inference results. (1,222 sample locations collected during 2000 - 2020 ) [ 27 ] . |
| GFM | Tree crop (palm, rubber, orchards) training samples (3,557 sample locations in 2015) [ 10 ] . |
| SERVIR Ucayali Palm | Annotated palm polygons in Peru (205 sample locations in 2015) [ 53 ] . |
| Palm Descals | Annotated palm polygons (66 sample locations in 2019) [ 14 ] . |
| MapBiomas | A collection of tree crop sources (can include manual annotations and inference results) in Brazil (14,971 sample locations collected during 2016 to 2019) [ 25 , 54 ] . |
| Palm Vollrath | Oil palm annotations (104 sample locations in 2017) [ 55 ] . |
Table A1 : Description of label sources used for the tree crop class. Also denoting the region (for regional dataset) and the number of sample locations.
| Name | Description |
|---|---|
| SDPT Planted forest | A collection of planted forest sources, including both, manual annotations and inference results [ 27 ] . |
| GFM | Naturally regenerating and planted/managed forest training samples [ 10 ] . |
| TMF | Tropical moist forest inference results [ 9 ] . |
| SBTN | A collection of natural and non-natural land [ 57 ] . |
| WorldCover | Inference map of land cover land use classes [ 58 ] . |
| Primary Humid Tropical Forests | Inference map of natural forests [ 59 ] . |
Table A2 : Label sources for the non-tree crop classes.
Figure A3 : Overview of the proposed method. The input consists of tensors from Sentinel-1 ( ) and Sentinel-2 ( ). The model generates a segmentation map of shape ( , is the number of classes). The dimensions handled by spatial encoder, temporal encoder and the decoder are highlighted in blue, purple, and gray, separately.
Figure A4 : Reliability diagram illustrating the relationship between predicted uncertainty and empirical accuracy. The purple line represents observed accuracy as a function of predicted uncertainty, while the dashed diagonal indicates perfect calibration. The Expected Calibration Error (ECE) is 8.1%.
Figure A5 : Overview of the uncertainty map
More in this topic
Agricultural Remote Sensing
Agricultural remote sensing uses satellite and airborne observations to map crops and fields and monitor yield, soil, and crop health.
See all 84 papers