HCL-Net: Heterogeneous Collaborative Learning for Lightweight Remote Sensing Image Segmentation

Semantic segmentation of remote sensing images remains challenging due to large intra-class variations, high inter-class similarity, and the demand for lightweight deployment. Conventional single-architecture models and homogeneous collaborative frameworks struggle to balance local detail extraction with global context modeling. To address these limitations, we propose HCL-Net, a heterogeneous collaborative learning framework that integrates convolutional and Transformer architectures. HCL-Net consists of two complementary student networks: the Frequency-domain Local Detail Network (FLDNet), based on ResNet18 with a wavelet phase–amplitude fusion block to capture multi-frequency information, and the Spatial-domain Global Structure Network (SGSNet), built on a DFormer-T backbone with a dynamic texture–edge perception module for robust global context modeling. A dual collaborative strategy enhances knowledge transfer between networks through (1) bidirectional feature reconstruction, which aligns high-order statistics using Gram matrix alignment and enforces feature-space consistency via variational information distillation, and (2) regional pixel-level contrastive learning, which improves intra-class compactness while reducing inter-class confusion. Experiments on the Vaihingen dataset demonstrate that collaborative training yields substantial gains over independent training, with FLDNet achieving mAcc 89.92% / mIoU 82.12% and SGSNet achieving mAcc 89.95% / mIoU 82.16%, improving accuracy by 2.26%/2.08% and IoU by 2.53%/2.19%, respectively. With only 24.25M and 12.36M parameters and computational costs of 6.13G and 6.35G, FLDNet and SGSNet outperform 19 state-of-the-art methods while remaining efficient for resource-constrained environments. Code and experimental results are available at https://github.com/110-011/HCL-Net.