A remote sensing image scene boundary identification method conforming to human visual perception (SBVP)

• SBVP delineates scene boundaries conforming to human visual perception. • SBVP uses image average state to guide precise boundary detection. • SBVP outperforms eCognition and SAM, achieving a 20–23 % improvement. • SBVP offers an automated, effective solution for scene boundary identification. Delineating clear scene boundaries in remote sensing images remains a challenge despite their ability to visually present landscapes. This study introduces a novel Scene Boundary Visual Perception (SBVP) method designed to identify scene boundaries in remote sensing images that conform to human visual perception. SBVP first computes the image’s average state as a fundamental unit for perception, guiding upscaling segmentation for initial scene targets and superpixel segmentation for object-level refinement. The hypothesis that the average state of the image scene can effectively guide upscaling segmentation and superpixel segmentation was tested on 16 large-scale remote sensing images with varying spatial resolutions and land cover types. SBVP was compared against eCognition and the Segment Anything Model (SAM) and further applied to suspicious target detection and the delineation of arid-humid and global biogeographical boundaries. Experimental results validate the hypothesis, showing that SBVP outperforms eCognition and SAM, achieving a mean IoU of 0.90, representing a 20 % and 23 % improvement over eCognition and SAM, respectively ( P < 0.01). Additionally, SBVP effectively detects suspicious targets and enhances the delineation of arid-humid and global biogeographical boundaries. As a simple yet effective approach, SBVP provides an automated solution for identifying scene boundaries that closely conform to human visual perception, offering broad applicability in remote sensing and geographic studies.