Entropy guidance hierarchical rich-scale feature network for remote sensing image semantic segmentation of high resolution

Abstract

Semantic segmentation of high-resolution remote sensing images (HRRSIs) is crucial for a wide range of applications, such as urban planning and disaster management. However, in HRRSIs, existing multiscale feature extraction and fusion methods often fail to achieve the desired accuracy because of the challenges posed by densely distributed small objects and large-scale variations. Therefore, we propose a hierarchical rich-sale feature network with entropy guidance (HRFNet), which introduces an entropy-based weighting and feature mining strategy to enhance feature extraction and fusion. Specifically, image entropy is employed as a quantifiable index to characterize the object distribution within remote sensing images, enabling an adaptive image division strategy. The image entropy is further used as weights during network training to emphasize regions with high entropy, which often correspond to edges and densely populated small objects. Additionally, the proposed feature mining strategy effectively integrates both global and local contextual information across multilayer feature maps. Extensive experiments show that HRFNet achieves mIoU scores of 81.31%, 86.47%, and 51.5% on the Vaihingen, Potsdam, and LoveDA datasets, respectively, outperforming existing methods by 1.0–3.0% mIoU.