Enhancing above-ground biomass estimation in agroforestry systems: A scalable deep learning approach using Sentinel-2 data

Above-Ground Biomass (AGB) is a critical indi2 cator for monitoring vegetation and carbon dynamics. This is 3 particularly true in agroforestry systems, where AGB reflects 4 the potential carbon sequestration of trees integrated into agri5 cultural landscapes, supporting climate change mitigation and 6 sustainable land management. In this study, we used publicly 7 available Sentinel-2 satellite data to map AGB in agroforestry 8 systems, where the discontinuous tree canopy, designed to allow 9 sunlight for understory crops, makes the spatial and temporal 10 resolution of the sensor particularly well suited to capture the 11 unique vegetation dynamics of this ecosystem. First, a Convolu12 tional Neural Network (CNN) enhanced by a Convolution Block 13 Attention Module (CBAM), is pre-trained with 30000 patches 14 using the high-resolution aerial Light Detection and Ranging 15 (LiDAR) acquired in the Netherlands as reference. The large16 scale pre-training enables the model to effectively capture key 17 features of Sentinel-2 for AGB estimation. The model is then 18 tested, with and without fine-tuning, in four different agroforestry 19 study areas in Africa and Latin America. Validation against ref20 erence airborne LiDAR data demonstrates strong generalization 21 ability of the proposed approach with R² values of 0.79, 0.73, 22 0.82 and 0.81 for Tanzania, Colombia, Nicaragua and Peru, 23 respectively. Our findings highlight the potential of combining 24 LiDAR-based structural information with deep transfer learning 25 to support large-scale biomass estimation and carbon monitoring 26 in smallholder agroforestry landscapes.