SAMOIS: efficient fine-tuned SAM with multi-scale enhancement for optical remote sensing image segmentation

Remote sensing image segmentation is a fundamental task in automated image analysis, aiming to assign accurate semantic labels to each pixel in sensor-acquired imagery. Although the Segment Anything Model (SAM) performs well on general image segmentation tasks, its effectiveness declines in non-natural domains like remote sensing due to limited semantic labels and poor boundary delineation. To overcome this limitation, we propose the SAMOIS framework. Based on the pre-trained SAM model, the framework improves the encoder’s adaptability to remote sensing imagery using low-rank adaptation, while also reducing trainable parameters. A cross-layer Transformer feature fusion module is introduced, utilizing multi-scale convolutions to enhance the representation of complex geographical patterns. Furthermore, a multi-scale feature pyramid decoder combined with a frequency-domain channel cooperative attention mechanism further improves segmentation accuracy. Experimental results show that SAMOIS achieves 55.22% mIoU and 73.49% OA on the LoveDA dataset, and 78.94% mIoU and 90.73% OA on the Potsdam dataset, significantly outperforming existing methods and demonstrating its effectiveness in complex Earth observation scenarios.