AMHCNet: an Adaptive Multi-Scale Heat Conduction Network for Remote Sensing Image Semantic Segmentation

Abstract

Remote sensing image semantic segmentation is challenging due to complex spatial structures. However, existing methods often struggle to fully exploit hierarchical contextual information, leading to limited segmentation performance. To alleviate this, we propose AMHCNet, an Adaptive Multi-Scale Heat Conduction Network, to enhance hierarchical feature representation and contextual propagation for remote sensing image semantic segmentation. The network contains two key components: a Scale-Specific Heat Conduction (SSHC) module that applies heat conduction with scale-dependent strength to each encoder stage, and an Adaptive Multi-Scale Feature Fusion (AMFF) module that dynamically fuses heat-enhanced features with adaptive weights. SSHC captures both local details and long-range semantic context, while AMFF preserves fine structures and aggregates global information. Extensive experiments on ISPRS Potsdam and Vaihingen datasets demonstrate that AMHCNet achieves better performance compared with several state-of-the-art methods while maintaining competitive computational efficiency.