Atlantic Forest gets wall-to-wall canopy height and carbon maps, with regional model error under half the global product's
Synopsis
The Brazilian Atlantic Forest now has 2024 wall-to-wall maps of canopy height at 10 m and aboveground carbon density at 30 m, with height predicted by a deep learning network fusing optical, radar and elevation data and trained and tested on airborne LiDAR from São Paulo state. On held-out test patches the 10 m height root mean square error is 3.05 m and the 30 m error is 2.58 m; over the same pixels the Meta and World Resources Institute global product has more than twice this error and underestimates canopy height by about a third.
Interpretation
The Brazilian Atlantic Forest now has wall-to-wall canopy height and aboveground carbon density maps for 2024, at 10 m and 30 m resolution respectively. The region previously lacked airborne LiDAR reference data sufficient to train a regional model, and large-scale carbon maps relied mostly on global products whose regional accuracy was rarely tested. On 1,793 held-out test patches in São Paulo state, the 10 m height root mean square error is 3.05 m and the 30 m error is 2.58 m; at two LiDAR sites outside São Paulo the error stays below 5 m.
The height prediction network encodes four satellite inputs in separate branches and fuses them before regressing height. Each sensor keeps a branch at its native resolution, and alignment to the 10 m optical grid happens only in feature space, avoiding upsampled radar that would invent spatial detail. Training used 7,179 patches and testing 1,793 patches, about 58,800 km2 in total, from a 2023–2024 statewide airborne LiDAR campaign over São Paulo state.
Canopy height is converted to aboveground carbon density through a power law fitted to 124 forest inventory plots. The carbon layer comes with a 30 m standard error and a 1 km summary of mean carbon density and its uncertainty. Canopy height explains 76% of the variance in plot carbon density; remaining forest across ten natural domains covers 28.3 Mha and holds 1,944 Tg C at a mean density of 68.8 Mg C ha-1.
Perspective
The layers describe the Brazilian Atlantic Forest in 2024, and height accuracy is assessed only on pixels inside the MapBiomas forest mask with LiDAR height above zero; training comes almost entirely from São Paulo state, with external validation at only two sites, in southern Bahia and Santa Catarina.
The height bias varies with height: canopy below 10 m is overestimated by about 1.25 m, the 30 to 40 m class is underestimated by 6.47 m, and canopy above 40 m by 16.38 m, and the carbon layer inherits and amplifies this compression. At Serra do Conduru in southern Bahia the bias is +2.49 m and grows with height, while the Santa Catarina site is covered by fine-tuning patches, so its error is not independent of the training data. The carbon standard error excludes the error of the predicted height and is therefore a lower bound.
