A Multi-Scale Boundary-Aware Deep Network for High-Resolution Remote Sensing Image Segmentation

Abstract

High-resolution remote sensing image segmentation plays an important role in land-cover mapping, urban monitoring, environmental assessment, agricultural analysis, and disaster management. However, this task remains challenging due to large scale variations, complex backgrounds, small structures, class imbalance, and ambiguous object boundaries. This paper proposes Multi-Scale Boundary-Aware MobileNet (MSBA-MobileNet), a multi-scale boundary-aware deep network for high-resolution remote sensing image segmentation. The proposed model combines a lightweight MobileNet-based feature extractor, explicit multi-level feature fusion, and a boundary-aware refinement branch. A hybrid training objective with weighted cross-entropy, Dice loss, and boundary supervision is used to improve both region-level segmentation and contour-level prediction. Experiments were conducted on a balanced subset of the LoveDA dataset, containing 1,500 training images and 400 validation images from urban and rural domains. The proposed MSBA-MobileNet achieved an overall accuracy of 0.6138, a mean Intersection over Union (mIoU) of 0.4253, a mean F1-score of 0.5917, and a Boundary-F1 score of 0.2914. Compared with TinyUNet, MSBA-MobileNet improved the mIoU by approximately 74.8%. Compared with LRASPP-MobileNetV3, the proposed model achieved competitive segmentation accuracy with approximately 61.8% fewer trainable parameters and the highest Boundary-F1 score among the evaluated models. The results indicate that multi-scale feature fusion and boundary-aware refinement are beneficial for compact remote sensing segmentation models.