Integrating global and local features for multisource remote sensing image classification with 2D Selective Scan Fusion

Multisource remote sensing image classification poses significant challenges due to the need to integrate spatial and spectral information from heterogeneous data sources such as HSI and LiDAR/SAR. Traditional methods, particularly Transformer-based models, often struggle to balance local feature extraction with long-range dependency modelling, leading to performance trade-offs. To address these limitations, we propose IGLNet (Integrating Global and Local Features Network), a novel framework with an efficient 2D Selective Scan Fusion mechanism. Specifically, the Spa-Conv-SSM block integrates convolutional layers with state-space models (SSM) to capture both local and global spatial structures, and is applied to HSI as well as LiDAR/SAR data for effective spatial representation learning with reduced complexity. In contrast, the Spe-Conv-SSM block is tailored to HSI data, exploiting its rich spectral information through the same convolution – SSM synergy. To reconcile modality heterogeneity, we introduce the Spatial Feature Fusion Module (SFFM), whose core 2D Selective Scan (SS2D) mechanism explicitly models cross-modal correlations. By selectively emphasizing spatial patterns that preserve both geometric and spectral cues, SFFM enables more accurate and robust feature fusion. Extensive experiments on the Berlin, Augsburg, and Houston2018 datasets demonstrate that IGLNet consistently outperforms state-of-the-art approaches in both accuracy and robustness, establishing its effectiveness for multisource remote sensing classification tasks.