GLFRNet: Global-Local Alignment and Frequency Reconstruction for Multimodal Remote Sensing Semantic Segmentation

High-precision semantic segmentation of multimodal remote sensing imagery (RSI) is critical for applications such as urban planning and land monitoring. However, existing methods often suffer from cross-modal semantic misalignment, geometric discontinuity, and loss of frequency information, resulting in blurred boundaries and incomplete object representations. To overcome these challenges, we present GLFRNet, a novel global-local alignment and frequency reconstruction fusion network for multimodal semantic segmentation. GLFRNet introduces three core innovations. First, a dual-stage structure-aware encoder (DSAE) is designed to combine convolutional and state space modeling, thereby enhancing both local texture extraction and global geometric feature representation. Second, a global-local feature projection alignment adapter (FPAA) is incorporated to explicitly align semantic and structural information between optical and digital surface model (DSM) modalities. Third, a phase-guided frequency reconstruction module (PGFRM) leverages the Fourier transform to jointly reconstruct low-frequency (LF) structures from DSM and high-frequency (HF) textures from optical images. Extensive experiments on the ISPRS Vaihingen and Potsdam datasets demonstrate that GLFRNet achieves state-of-the-art performance, with mean intersection-over-union (mIoU), overall accuracy (OA), and mean F1 -score (mF1 ) scores of 87.03%, 91.93%, and 92.83% on Potsdam, and 85.35%, 93.27%, and 92.33% on Vaihingen, respectively. GLFRNet significantly outperforms previous methods, particularly in segmenting elongated objects, occluded regions, and complex terrain. All codes will be available on https://github.com/darkseid-arch/GLFRNet