Remote Sensing Semantic Change Detection Based on the Visual Foundation Model

Semantic change detection (SCD) supplies not only the change location but also the detailed land cover/land use (LC/LU) categories before and after the change. Vision Foundation Models (VFMs) such as the Segment Anything Model (SAM) allow for zero-shot or reduce the reliance on task-specific modeling expertise, making it possible for them to migrate to new image distributions and tasks. However, due to the special imaging characteristics of Remote Sensing (RS), their direct use in many remote sensing applications is often unsatisfactory. Therefore, our goal is to utilize the powerful visual recognition capability of VFM to improve semantic change detection in high-resolution remote sensing images (RSIs). To make MobileSAM applicable to RS scenarios, we fine-tune its encoder.