Knowledge-Aware Progressive Fusion Network for Heterogeneous Remote Sensing Image Semantic Segmentation

The advancement of remote sensing technology has enabled multimodal semantic segmentation using optical image and synthetic aperture radar data. However, two critical challenges limit existing fusion methods: modal shift, where mismatched feature distributions cause network overfitting to the dominant modality; and data heterogeneity, reflected in inconsistencies from distinct imaging mechanisms, including noise characteristics, frequency-domain features, and semantic representation discrepancies. To address these challenges, we propose a knowledge-aware progressive fusion network (KPFNet) with three key innovations. First, a two-stage learning strategy with self-supervised feature decoupling independently trains modality-specific branches under intermodal correlation constraints, mitigating modal shift and preventing single-modality dominance. Second, the progressive feature fusion strategy adopts a “correction-then-fusion” approach, utilizing differential calibration which involves strong spatial-channel calibration for high-frequency features and weak channel calibration for low-frequency features, in order to align feature details and global structures. Third, a semantic knowledge-guided layer injects semantic priors to reduce intraclass discrepancies and enhance interclass separability, overcoming semantic ambiguities. Experiments on two public datasets demonstrate that KPFNet achieves superior performance compared to state-of-the-art methods.