Study on Segmentation Potential of SAM on Medium Resolution Remote Sensing Images

Segmentation of remote sensing imagery plays a crucial role in Earth observation, supporting various domains such as crop surveillance, ecological evaluation, and land classification. In recent years, the rapid advancement of large-scale vision models has significantly improved remote sensing image processing techniques, achieving notable success in high-resolution imagery. However, their segmentation performance on medium-resolution remote sensing images remains underexplored. This study employs the Segment Anything Model (SAM) to extract black shade net features from Sentinel-2 imagery. Segmentation accuracy is assessed using the COCO evaluation metric. Experimental results demonstrate that the model effectively identifies large shade net regions, achieving a mean Average Precision (mAP) of 80%. For medium- and small-sized targets, the mAP values are 52.7% and 32.7%, respectively. While the model exhibits strong segmentation performance for large-scale features in complex backgrounds, its ability to detect smaller targets requires further improvement.