Reconstructing Multi-Decadal Forest Disturbances: A Spatio-Temporal Transformer Approach
Researchers have developed a deep learning framework using a vision transformer to map forest disturbances across the contiguous United States over a 38-year period. This approach simultaneously models temporal trajectories and spatial neighborhoods, offering a more coherent alternative to traditional pixel-wise analysis of satellite data. Evaluations using multiple satellite sources and validation datasets show high precision in detecting disturbances, though performance varies across different disturbance types compared to existing methods. AI
IMPACT This method offers a more spatially coherent approach to monitoring environmental changes, potentially improving land management and carbon dynamics analysis.