Wall-to-wall maps of canopy height and aboveground carbon across the Atlantic Forest
Authors: Matheus Pinheiro Ferreira, Thiago Carmuega Rabaçal, Giulio Brossi Santoro, Giovanna de Andrade Ferreira, Rodrigo Costa Pinto, Danilo R. A. de Almeida, Paulo Guilherme Molin, Catherine Torres de Almeida, Angélica F. Resende, Vinicius Castro Souza, Marcelo Antonio Pinho Ferreira, Cássio Toledo, Eben North Broadbent, Angelica Almeyda Zambrano, Laura Simões, Juliano van Melis, Ricardo Ribeiro Rodrigues, Pedro H. S. Brancalion
Organizations: Department of Forest Sciences, “Luiz de Queiroz” College of Agriculture, University of São Paulo (USP/ESALQ), Piracicaba, SP, Brazil · Bioflore Technology, Brazil · Center for Carbon Research in Tropical Agriculture (CCARBON), University of São Paulo (USP), Piracicaba-SP, Brazil · Instituto Geográfico e Cartográfico de São Paulo (IGC), São Paulo, SP, Brazil · Center for Nature Sciences, Federal University of São Carlos (UFSCar), Buri, SP, Brazil · Department of Geography, Federal University of Paraná (UFPR), Curitiba, PR, Brazil · Department of Biological Sciences, “Luiz de Queiroz” College of Agriculture, University of São Paulo (USP/ESALQ), Piracicaba, SP, Brazil · Spatial Ecology and Conservation (SPEC) Lab, School of Forest, Fisheries, and Geomatics Sciences, University of Florida, Gainesville, FL 32611 USA · Universidade Federal de São Carlos, Araras, Brazil · Re.green, Praça Santos Dumont, 70 – 4° Andar, Rio de Janeiro-RJ, Brazil
Global canopy height maps are increasingly used to estimate forest carbon, yet their regional accuracy is rarely tested. We present wall-to-wall maps of canopy height (10 m) and aboveground carbon density (30 m) for the Brazilian Atlantic Forest in 2024. Height was predicted by a deep learning network fusing Sentinel-2 reflectance, Sentinel-1 and ALOS PALSAR backscatter and ALOS elevation, trained on 7,179 and tested on 1,793 patches (58,800 km2) from a statewide airborne LiDAR campaign over São Paulo state, Brazil. On held-out test patches, the mean absolute error is 2.26 m at 10 m and 1.90 m at 30 m, and the model was validated at two LiDAR sites outside São Paulo. Over the same pixels, the Meta and World Resources Institute global product has more than twice this error. A power law fitted to 124 inventory plots converted canopy height to above ground carbon density. The remaining forest holds 1,944 TgC over 28.3 Mha. Global products therefore need local validation before use in carbon accounting.
Figures & tables
Figure 1: Study area. The Atlantic Forest region is outlined in the inset; the main panel shows the State of São Paulo, the reference window from which the training and test patches of the calibration grid were drawn and which contains a variety of forest formations. Cyan and magenta polygons indicate, respectively, the training and testing patches of size 2.56 × 2.56 km. Red dots mark the 124 field plots used to convert canopy height into aboveground carbon density.
Figure 2: Examples of training patches from four contrasting landscapes, (a) continuous montane forest, (b) forest–pasture mosaic, (c) forest plantation and (d) urban–forest interface. Every row covers the same 2,560 × 2,560 m ground footprint, and each source is displayed on its own native grid, without resampling (Table 1 ). Columns, from left to right, show Sentinel-2 true-colour composite (R: B4, G: B3, B: B2; 10 m); Sentinel-1 false-colour composite (R: VV, G: VH, B: VV/VH ratio, in dB; 10 m); ALOS PALSAR false-colour composite (R: HH, G: HV, B: HH/HV ratio, in dB; 25 m); ALOS DSM rendered as shaded relief (30 m), with the elevation range of the patch in the lower-right corner; MapBiomas 2024 forest mask (“MapBiomas forest mask”), with the percentage of forest pixels in the lower-right corner; and the LiDAR-derived reference CHM (10 m). Sentinel-2 and the DSM are contrast-stretched per patch, whereas each SAR band uses a single 2–98% percentile stretch shared by the four patches, so that radar tones are comparable across rows. Heights above 40 m are shown with the top colour of the CHM scale. The scale bar in the Sentinel-2 column applies to all panels. In (c), tall planted stands are mapped as non-forest by MapBiomas: they are used in training but excluded from the accuracy assessment (“Accuracy assessment”). In (a) and (b), the slope-driven brightness pattern in both SAR composites illustrates the topographic effect that motivates the DSM branch (“ALOS DSM (elevation)”).
Figure 3: Four-branch fusion architecture for canopy height estimation. Each modality is processed at native resolution: Sentinel-2 via BPAT-UNet (incorporating boundary enhancement [BEM] and channel attention [CA]), while Sentinel-1, ALOS PALSAR, and ALOS DSM use lightweight U-Nets. Branch outputs (64 channels each) are bilinearly aligned to the 256 × 256 Sentinel-2 reference grid. A gated cross-modal fusion module combines the 256 channels, and a regression head predicts the final canopy height map. BEM, CA, and fusion blocks are detailed in Fig. 4 .
Figure 4: Building blocks of the fusion network (Fig. 3 ). Top. BPAT-UNet’s boundary enhancement module (BEM) and channel attention (CA). BEM extracts a single-channel boundary map B to gate a residual enhancement E(X) added to the input. CA then reweights the channels of X′ using pooled descriptors and a shared MLP. Bottom. Gated cross-modal fusion. Concatenated features Fcat are globally pooled and fed to a bottleneck MLP to generate a 64-channel gate per modality. Modality features are channel-wise scaled by their respective gates, re-concatenated, and refined via 3×3 convolution before the regression head.
Figure 5: Distribution of LiDAR canopy height in the São Paulo patches and in the two external validation sites. Each band is a histogram in classes of 1 m built over the pixels that fall inside the MapBiomas forest mask and carry a height above zero. The interquartile range is shaded and the median is marked, and the median, the 95th percentile and the number of pixels are reported on the right of each band. The vertical bar gives the share of pixels falling in a class of 1 m, the horizontal axis is truncated at 35 m and the bands are ordered by median height. The forest mask is what makes the comparison one of canopy rather than of land cover.
Figure 6: Canopy height and aboveground carbon density of the Brazilian Atlantic Forest in 2024. (a) Canopy height (m) and (b) aboveground carbon density of record 2, both drawn over the Atlantic Forest region, filled in black, in an Albers equal-area projection. The insets show the same layers close to their native resolution at (c, e) the coast of Bertioga, in São Paulo, where the tall forest of the Serra do Mar escarpment gives way to the lower restinga and mangrove forests of the coastal plain, and (d, f) the Estação Veracel, a private reserve of lowland rainforest in southern Bahia.
Figure 7: Systematic error of the four-branch model (E5) by reference canopy height derived from LiDAR. Bias is the mean of LiDAR minus predicted height, so bars above zero, in red, denote underestimation and bars below zero, in blue, denote overestimation. The right-hand axis gives the same bias as a percentage of the class mean height, the label above or below each bar repeats that percentage together with the class RMSE, diamonds mark the median residual and error bars span ± 1 SD of the per-pixel residual within the class. The lower panel gives the number of test pixels in each class and its share of the total.
Figure 8: Prediction versus LiDAR canopy height for the four-branch model (E5) at (a) the native 10 m resolution and (b) 30 m, after aggregating both the prediction and the reference by spatial averaging. The dashed line is the 1 to 1 line and the solid line is the ordinary least-squares fit, with bias defined as LiDAR − predicted. Hexagon colour is the number of pixels per hexagon on a logarithmic scale shared by both panels and truncated at the 99th percentile (pointed end of the colour bar). Aggregation lowers RMSE and MAE and leaves the bias unchanged, which indicates that the error it removes is spatially uncorrelated rather than a systematic offset. The values are given in the text.
Figure 9: External validation of the four-branch model (E5). LiDAR canopy height against predicted canopy height at (a) Serra do Conduru, in southern Bahia, over 454,849 pixels, and (b) Jaraguá do Sul, in Santa Catarina, over 51,331 pixels. Hexagon colour is the number of pixels per hexagon, on a scale of its own in each panel, the dashed line is the 1:1 line and the solid line is the ordinary least-squares fit of reference on prediction. The two panels share the same axes, from 0 to 50 m on both.
Figure 10: Comparison with the global canopy height product of Meta and WRI over the test patches of São Paulo. Panel (a) is the four-branch model and panel (b) is the external product, both against the same LiDAR reference and over the same 53,755,301 pixels, restricted to the MapBiomas forest mask. Hexagon colour is the number of pixels per hexagon, the dashed line is the 1 to 1 line and the solid line is the linear fit.
Figure 11: Aboveground carbon density (AGCD) as a function of top canopy height (TCH) for the 124 inventory plots, with the plots coloured by category. The solid line is Equation 12 with a=0.089 and b=1.645 , and the box gives these coefficients together with the goodness-of-fit statistics evaluated on the plots, all in kg m -2 , where 1 kg m -2 equals 10 Mg C ha -1 . Bias is the mean of observed minus predicted carbon density, so a positive value means that the curve lies below the plots. The number of plots in each category is given in the legend.
Figure 12: Standard error of the aboveground carbon density (AGCD) map. (a) Standard error of the 30 m pixels (Equation 14 , record 3), averaged over the forest pixels of display cells of about 1.2 km. Grey marks the parts of the region where forest covers less than 3% of the cell. (b) Standard error of a pixel as a function of canopy height, solid within the height range of the 124 plots and dashed beyond the tallest plot (24.0 m), with dots giving the relative standard error at 5, 10 and 20 m. The error of the predicted height is not included.
Figure 13: Aboveground carbon density of the remaining forest in the ten natural domains of the Atlantic Forest. (a) The domains coloured by their mean AGCD, with the part of the region that the domain layer does not cover shown in grey. (b) Area-weighted distribution of the 30 m AGCD inside each domain, in classes of 2 Mg C ha -1 and ordered by median, with the interquartile range in the darker shade and the median as a vertical line, followed by the forest area of the domain and by its total carbon stock. Colour carries the mean AGCD on the same scale in both panels, and the numbers on the map follow the order of the bands.
World Resources Institute, 10 G St NE #800, Washington, DC 20002, USA · Fundamental AI Research (FAIR), Meta, 75002 Paris, France · Meta, 1 Hacker Way, Menlo Park, CA 94025, USA +1
University of Alberta, Alberta Centre for Earth Observation Sciences, Department of Earth and Atmospheric Sciences, University of Alberta, AB, T6G 2E3, Edmonton, Canada · Universitat Politècnica de València, Geo-Environmental Cartography and Remote Sensing Group (CGAT), Department of Cartographic Engineering, Geodesy and Photogrammetry, Camí de Vera s/n, 46022, València, Spain · University of Los Andes, Photogrammetry and Remote Sensing Laboratory, Vía Chorros de Milla, Mérida, 1511, Mérida, Venezuela
Department of Information Systems, University of Münster, Germany · Laboratoire des Sciences du Climat et de l’Environnement (LSCE), France · Zuse Institute Berlin (ZIB), Germany +2