TCCFNet: a semantic segmentation method for mangrove remote sensing images based on two-channel cross-fusion networks

Lixiang Fu ¹ Yaoru Wang ¹ Shulei Wu ^1* Jiashen Zhuang ¹ Zhongqiang Wu ¹ Jian Wu ² Huandong Chen ¹

• ¹ Hainan Normal University, Haikou, China
• ² China Mobile Group Hainan Company Limited, HaiKou, China

Mangrove ecosystems play vital roles in coastal environments. Deep learning-based image segmentation has emerged as a crucial tool for mangrove research and monitoring.However, current methods face significant challenges in distinguishing similar categories and effectively capturing information from large target objects in mangrove remote sensing images.To address these limitations, we propose TCCFNet (Two-Channel Cross-Fusion Network), a semantic segmentation method that leverages a dual backbone architecture. The network combines ResNet for detailed local feature extraction with Swin Transformer for comprehensive global context modeling. This complementary structure enables robust feature representation across different scales and contexts.To strengthen feature integration, we introduce the Cross Integration Module (CIM). This module enhances the network's feature processing capabilities through a dual-path design. The first path employs a cross-attention mechanism to highlight significant features, while the second path implements a multi-scale feature fusion strategy to combine information effectively across different levels of the network.The experimental results show that TCCFNet has excellent performance in mangrove remote sensing image segmentation, the average cross-linking degree (MIoU) is 88.34%, the pixel accuracy (PA) is 97.35%, and the F1-score is 93.55%.Notably, the model demonstrated excellent ability in handling challenging scenarios, especially in distinguishing similar categories and accurately segmenting large target areas.Compared with existing methods such as MSFANet (86.02%) and DC-Swin (86.26%), TCCFNet has significantly improved segmentation accuracy, especially in complex mangrove environments, where accurate boundary detection is crucial.