DeepForest: Novel Deep Learning Models for Land Use and Land Cover Classification Using Multi-Temporal and -Modal Sentinel Data of the Amazon Basin

Land use and land cover (LULC) mapping is a powerful tool for monitoring large areas. For the Amazon rainforest, automated of critical importance, as changing rapidly due to forest degradation deforestation. Several research groups have addressed this challenge by conducting local surveys producing maps using freely available remote sensing data. However, automating process large-scale remains one biggest challenges in community. One issue when supervised learning scarcity labeled training way address problem make already produced with (semi-) classifiers. This also known weakly learning. The present study aims develop novel methods LULC classification cloud-prone basin (Brazil) based on labels from MapBiomas project, which include twelve classes. We investigate different fusion techniques multi-spectral Sentinel-2 data synthetic aperture radar Sentinel-1 time-series 2018. newly designed deep architectures—DeepForest-1 DeepForest-2—utilize spatiotemporal characteristics, well multi-scale representations In several scenarios, models are compared state-of-the-art (SotA) models, such U-Net DeepLab. proposed networks reach an overall accuracy up 75.0%, similar SotA models. approaches outperform respect underrepresented Forest, savanna crop were mapped best, F1 scores 85.0% combining multi-modal data, 81.6% reached Furthermore, qualitative analysis, we highlight that classifiers sometimes inaccurate labels.