Due to the significant intra-class variation, small target detection problem under low-contrast clutter background, accurate semantic segmentation of very high resolution remote sensing images still exists many difficulties at present. For the above problems, we introduce SAM2-MAFLNet as an improved teacher-student knowledge-distillation approach to build a low-resource model that achieves high-density segmentation of remote sensing images. The student develops a Multi-attention and cross-fusion architecture (MACA) that jointly learns long-range spatial relationships and channels dependency, and an enhanced Laplacian high-frequency enhancement module (LHFEM) introduces edge-aware information via laplace pyramid decomposition and foreground-background segmentation. Experiments on the ISPRS Vaihingen and Potsdam benchmarks show that the teacher network achieves 91.92% mF1/85.06% mIoU and 93.64% mF1/88.21% mIoU, respectively, and the student retains competitive accuracy with only 12.52M parameters. The above experiments show that, in terms of balancing segmentation accuracy and real-time requirements for application on remote-sensing data.
