Large Structure Change Detection in Medium-Resolution Satellite Imagery Via Transfer Learning And Scaling Strategies

This study addresses the deficiency in change detection methods for large structures within medium-resolution (MR) satellite imagery, focusing on the specific case of detecting building alterations. Adapting the successful CSA-CDGAN [6] framework originally designed for high-resolution imagery to the unique challenges of 5m resolution, we explore its efficacy in detecting changes across diverse landcovers and seasonal variations. Utilizing datasets created from VENμS satellite MR imagery, we conduct experiments demonstrating model performance improvements through transfer learning and image scaling strategies. The results showcase adaptability not only to seasonal changes but also to 10m Sentinel-2 imagery, emphasizing the approach’s robustness in detecting subtle alterations across varying medium spatial resolutions. Our findings highlight the potential of advanced deep learning techniques in enhancing change detection capabilities for MR satellite imagery, particularly in scenarios with extended temporal spans and complex environmental transformations.