A Comparative Deep Learning Approach to Battle Damage Detection with Foundation Model Architectures and Open Access Sentinel-1 SAR Data

Building damage assessment is vital for humanitarian responses, yet timely SAR data presents challenges like speckle and class imbalance. Despite advances in Earth Observation Foundation Models(EO-FM), they remain largely poorly validated for distribution-sparse and challenging use cases like battle damage detection. This study addresses this gap by evaluating EO-FM for Sentinel-1 SAR battle damage assessment against statistical and deep learning baselines. Using a multi-city dataset, we compared U-Nets against Copernicus-FM, testing composite losses and Siamese architectures. Results show that while foundation models offer generalisability, they underperform on rare-class detection compared to Siamese U-Nets. Siamese U-Net Classifiers achieved an F1 of 0.81 versus Copernicus-FM’s 0.14. This shows that foundation models require task-aligned pretraining and tokenisation for automated precision monitoring.