LSLFormer: A Lightweight Spectral–LiDAR Fusion Network for Remote Sensing Image Classification

The fusion of hyperspectral images (HSI) and LiDAR data provides rich complementary spectral and elevation information for land cover classification. However, existing fusion methods, particularly Transformer-based models, are often constrained by high computational costs and complex cross-modal interaction mechanisms. To address this challenge, we propose a Lightweight Spectral-LiDAR Fusion Network (LSLFormer), which aims to achieve efficient and accurate remote sensing image classification. We introduce three key modules into LSLFormer architecture: 1) A Hyperspectral-to-Multispectral (H2M) module to alleviate the computational burden of self-attention on high-dimensional spectral data. 2) A Multi-scale Channel Interaction Enhancement (MCIE) module to extract robust spatial-structural features from LiDAR data. 3) A Spectral-LiDAR Attention (SLA) module to achieve deep cross-modal interaction by dynamically fusing normalized spectral and structural affinity distributions. In addition, we also propose a novel Cross-Modal Structural Consistency (CMSC) loss. This mechanism aligns the geometric topology of the spectral features with the LiDAR structure via knowledge distillation, ensuring precise boundary delineation without compromising spectral semantics. Extensive experiments on three public benchmark datasets have demonstrated that LSLFormer consistently outperforms other state-of-the-art convolutional and Transformer-based methods in terms of classification accuracy and computational cost. The codes of this work will be available at https://github.com/DianLi2002/LSLFormer.