Mapping the Brazilian savanna’s natural vegetation: A SAR-optical uncertainty-aware deep learning approach

The Brazilian savanna (Cerrado) is considered a hotspot for conservation. Despite its environmental and social importance, the biome has suffered a rapid transformation process due to human activities. Mapping and monitoring the remaining vegetation is essential to guide public policies for biodiversity conservation. However, accurately mapping the Cerrado’s vegetation is still an open challenge. Its diverse but spectrally similar physiognomies are a source of confusion for state-of-the-art (SOTA) methods. This study proposes a deep learning model to map the natural vegetation of the Cerrado at the regional to biome level, fusing Synthetic Aperture Radar (SAR) and optical data. The proposed model is designed to deal with uncertainties caused by the different resolutions of the input Sentinel-1/2 images (10 m) and the reference data, derived from Landsat images (30 m). We designed a multi-resolution label-propagation (MRLP) module that infers maps at both resolutions and uses the class scores from the 30 m output as features for the 10 m classification layer. We train the model with the proposed calibrated dual focal loss function in a 2-stage hierarchical manner. Our results reached an overall accuracy of 70.37%, representing an increase of 15.64% compared to a SOTA random forest (RF) model. Moreover, we propose an uncertainty quantification method, which has shown to be useful not only in validating the model, but also in highlighting areas of label noise in the reference. The developed codes and dataset are available on Github . • We investigated SAR-Optical deep learning data fusion to map the natural vegetation cover classes in the Brazilian Savanna. • We use Sentinel-1 and 2 imagery as input to learn from a Landsat-based reference map. • We propose the multi-resolution label-propagation module and the calibrated dual focal loss function to handle noisy labels. • Uncertainty evaluation indicates that the noisy reference underestimates the accuracy of our results.