Remote sensing image classification and visual recognition based on multimodal information fusion

The integration of heterogeneous remote sensing data from multiple sensors—such as optical, SAR, hyperspectral, and LiDAR—offers rich spatial and spectral information but also poses significant challenges in feature alignment and fusion. To address these issues, this paper proposes a novel Multimodal Information Fusion Network (MIFNet) for remote sensing image classification and visual recognition. The framework introduces a Cross-Modal Attention Fusion (CMAF) module that dynamically learns inter-modality dependencies, allowing adaptive weighting of spectral, structural, and elevation features. An Adaptive Alignment Module (AAM) further employs contrastive learning to reduce feature discrepancies between modalities and ensure coherent representation in latent space. Additionally, a multimodal saliency visualization technique enhances interpretability by revealing modality-specific contributions to classification outcomes. Experiments on Houston2018 and BigEarthNet datasets demonstrate that MIFNet achieves superior performance with up to 97.1% overall accuracy and 0.97 F1-score, surpassing state-of-the-art fusion models. The results confirm that MIFNet effectively enhances both classification accuracy and explainability in multimodal remote sensing analysis.