Speaker: Gabriel Belouze

Abstract: Forests play a central role in carbon sequestration, yet their contribution is difficult to measure due to the spatial and temporal heterogeneity of ecosystems.  Recent satellite missions provide free, regular coverage of the globe with medium- to high- resolution optical and radar imagery. In parallel, airborne and spaceborne LiDARs, as well as ground forest inventories, supply accurate but spatially sparse reference data on forest structure and biomass. Deep learning has proven effective for learning dense representations from these complementary sources. In this talk, I will present our ongoing work in canopy height and biomass mapping across the dry tropical forests of Tanzania and Mozambique. We leverage existing products to filter and correct GEDI LiDAR labels, which are noisy in this region, and train a convolutional network to predict height and biomass from Sentinel-1 (radar) and Sentinel-2 (optical) imagery. Early results shows improved estimation of tall canopy and provide the first high-resolution biomass maps for Tanzania and Mozambique.

Short Bio: Gabriel graduated from École Normale Supérieure (ENS) Paris with a degree in Computer Science and earned a master’s in Computer Vision and Machine Learning from the MVA program at ENS Paris-Saclay. Since 2023, he has been working at the Climate and Environmental Science Laboratory (LSCE) on canopy height mapping, continuing this work as a PhD candidate from late 2024 under the supervision of Fabian Gieseke (University of Münster) and Philippe Ciais (LSCE). He is currently starting my second year of PhD. His research focuses on computer vision for remote sensing applied to forest monitoring. Gabriel also enjoys functional programming, open-source collaboration, science openness and outreach, chess, and outdoor activities.