StructFuse-Net: A Structure-Aware Multimodal Fusion Network for Geometry-Consistent Optical–SAR Image Segmentation

Multimodal remote sensing segmentation remains challenging due to modality discrepancies, structural inconsistencies, and sensor-induced noise, which often lead to blurred boundaries and geometric distortions. To address these issues, we propose StructFuse-Net, a structure-aware multimodal fusion framework for geometry-consistent optical–SAR image segmentation. The framework comprises three complementary modules: (1) a Structure-Driven Directional Fusion (SDDF) module that enhances low-level representations by suppressing speckle noise and reinforcing geometric cues through cross-modal attention and wavelet-based directional decomposition; (2) a Semantic Refinement and Alignment Fusion (SRAF) module that improves high-level consistency by refining channel–spatial responses and aligning cross-modal semantics via residual guidance and frequency-aware enhancement; and (3) a Structure-Aware Cross-Level Fusion (SACLF) module that connects low- and high-level features through adaptive gating and orthogonal directional encoding, ensuring multi-scale geometry-consistent feature representation. Extensive experiments on the WHU-OPT-SAR and YESeg-OPT-SAR datasets demonstrate that StructFuse- Net consistently outperforms state-of-the-art baselines, achieving overall accuracies of 82.7% and 81.5% with mean IoUs of 54.5% and 62.4%, respectively. Moreover, experiments on a cloud-disturbed subset of the Pohang dataset further validate the robustness of StructFuse-Net under severe optical degradation, where it maintains coherent geometry and boundary continuity with smaller performance degradation than competing methods. Beyond quantitative improvements, the proposed network delivers superior boundary fidelity and spatial coherence, particularly for elongated and structure-sensitive objects such as roads and rivers. These results indicate that StructFuse-Net provides a robust and generalizable solution for geometry-consistent optical–SAR semantic segmentation.