MFTriNet: A Multimodal Fusion Triple-Flow Network for Remote Sensing Image Semantic Segmentation
Yan Wang, Li Gang Cao, He Deng · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · 2026
The fusion of high-resolution remote sensing imagery (HRRSI) with digital surface models (DSM) provides complementary spectral and height information, significantly enhancing the potential for accurate semantic segmentation. However, effectively reconciling the inherent domain disparities between these modalities and mitigating noise from defective data sources remain challenging. While prevailing encoder-decoder networks facilitate cross-modal fusion, their performance often degrades in complex scenes due to limited robustness to inter modal discrepancies and noise. To address these challenges, we propose MFTriNet, a novel multi-modal fusion triple-flow network that jointly considers modality-specific characteristics and cross-modal interactions. MFTriNet decomposes the segmentation task into two modality-specific flows and one modality complementary flow. The two modality-specific flows, sharing identical encoder-decoder architectures, independently process HRRSI and DSM data to generate preliminary segmentation predictions and preserve unique modal characteristics. In the modality-complementary flow, each encoding layer incorporates a dedicated modality-induced feature modulator (MFM) designed for feature alignment and discrepancy suppression, utilizing a strategy of correction followed by fusion, and each decoding layer integrates a modality-aware dynamic aggregation module (MDAM) for dynamic and gated attention-based fusion. The final prediction is generated by a learnable weighted fusion of outputs from all three flows. Extensive experiments on the ISPRS Vaihingen and Potsdam benchmarks demonstrate that MFTriNet achieves state-of-the-art performance, with mIoU/OA values of 84.21%/92.05% and 86.78%/91.78%, respectively. Furthermore, evaluations on the more challenging MMHunan dataset verify the effectiveness and generalization capability of MFTriNet in complex scenes. The source code is available at: https://github. com/YanWang-WHPU/MFTriNet.