Direct Segment Anything Model-Remote Sensing: A Vision-language Foundation Model for Semantic Edge Detection in Remote Sensing Imagery.

This article focuses on DirectSAM-remote sensing (DirectSAM-RS), a vision-language foundation model developed for semantic edge detection in remote sensing imagery. Built upon the Direct Segment Anything Model (DirectSAM) pre-trained on the SA-1B dataset, DirectSAM-RS integrates a large-scale RemoteContour-34k dataset of over 34,000 image-text-edge triplets and a prompter module that enables flexible conditioning on textual prompts for class-specific edge detection. Evaluated on multiple downstream benchmarks for coastline, building, and road extraction, DirectSAM-RS achieves state-of-the-art results in both zero-shot and fine-tuning scenarios, demonstrating improved generalization and efficiency compared to prior models. The study highlights the model's ability to unify multi-task learning across diverse semantic targets and suggests future work in expanding dataset size and exploring few-shot learning for remote sensing applications.