Hybrid Feature Extraction and Mamba Fusion for Multimodal Remote Sensing Image Classification

Multi-modal remote sensing image classification has emerged as a research focus. Mamba, characterized by linear computational complexity, has demonstrated potential in remote sensing image classification. However, when processing remote sensing images, the sequential modeling mechanism of Mamba inherently sacrifices fine-grained spatial details despite its ability to effectively capture global dependencies. This problem is further exacerbated when extracting high-level semantic features in deep architectures. To address these issues, we propose a hybrid feature extraction and multi-modal Mamba fusion (HFEMamba) framework. First, a Mamba-neighborhood attention synergistic extraction module (MASEM) is designed, which integrates Mamba with neighborhood attention through a parallel architecture to simultaneously capture global semantic information and local spatial details. Second, modality-specific feature enhancement modules (MSFE) are constructed to enhance the spectral discriminability of HSI and the spatial structure of LiDAR, thereby recovering fine-grained details lost during high-level feature extraction and preserving modality-specific features. Finally, an inter-modality Mamba interaction and fusion module (IMMIF) is proposed to facilitate multi-modal feature interaction and deep integration, effectively leveraging complementary information across modalities. Experimental validation on the Trento and Augsburg datasets demonstrates the superior performance of HFEMamba.