LGCFNet: A Feature Fusion Model Based on Larg-Kernel Convolution for Multimodal Remote Sensing Image Semantic Segmentation

Segmentation of high-resolution multispectral images is crucial for applications. Traditional methods that rely on single data sources face challenges like illumination variation, cloud cover, shadows, and similar spectral features, especially in segmenting objects of varying sizes. We propose a Large-Kernel Convolution-based Human-Eye Pyramid Cross-Modal Fusion Network (LGCFNet). In addition to RGB data, it fuses Digital Surface Model (DSM) information. In the encoder, the improved large-kernel convolutional network with light/heavy branches extracts features, and the Hybrid Feature Fusion Module integrates them. In the decoder, the Differential Feature Weighting Module adaptively aggregates feature information. Experiments on Potsdam datasets show LGCFNet improves segmentation performance by at least 2% in mIoU compared to state-of-the-art multimodal fusion models.