FreqFusion-Net: frequency-aware adaptive multimodal fusion for building extraction in complex terrain via transfer learning

Despite significant progress, deep learning-based building extraction from remote sensing still faces domain shifts across regional scenes and insufficient information from single data sources in complex environments. To address these issues, we propose FreqFusion-Net, integrating DWT-based frequency decomposition with the SegNeXt backbone. RGB imagery from Google Earth, NIR from Sentinel-2, and SAR from Sentinel-1 are decomposed into low-frequency structures and high-frequency details. The frequency-aware low-frequency fusion and scene-adaptive high-frequency gating modules enable spatially adaptive, modality-selective fusion, using a low-frequency context to dynamically compute per-modality reliability maps, selecting SAR in shadows, NIR in vegetation, or RGB in well-lit areas. We construct a mountainous multimodal dataset and apply transfer learning. The experiments yield an IoU of 0.616, surpassing SegFormer (0.577). Ablations validate each frequency component. Transfer learning boosts from zero-shot 0.313–0.616, confirming regional data's critical role and the superiority of frequency-aware fusion over conventional concatenation, offering an efficient solution for interpreting multi-source data in complex terrains.