Lightweight and adaptive multimodal fusion network for multisource remote sensing

Near-space remote sensing (NSRS) offers unique value for terrestrial surface monitoring and environmental perception due to its inherent capabilities for wide-area coverage and high spatial resolution. However, the heterogeneous nature of spectral and structural characteristics between hyperspectral images (HSI) and light detection and ranging (LiDAR) data has long impeded effective multimodal fusion and efficient classification, often leading to prohibitive computational costs and limited model adaptability. This paper proposes an adaptive multimodal attention convolutional network (AMACNet), which enables efficient integration of spectral and geometric information through a lightweight convolutional structure combined with a cross-modal global attention mechanism. Additionally, an adaptive complexity control (ACC) module is incorporated to dynamically align model capacity with dataset characteristics. Comprehensive experiments on three multimodal datasets validate the effectiveness of AMACNet. The results demonstrate that AMACNet surpasses state-of-the-art methods in both classification accuracy and computational efficiency, achieving a maximum overall accuracy (OA) of 88.4%, while reducing memory usage by 87.5% and increasing inference throughput sixfold.