Hybrid Deep Learning Model for Multi-Source Remote Sensing Data Fusion: Integrating DenseNet and Swin Transformer for Spatial Alignment and Feature Extraction

The integration of multi-source remote sensing data, including Synthetic Aperture Radar (SAR), optical, and hyperspectral imagery, is critical for enhancing Earth Observation Systems but is challenged by highdimensional variability and spatial misalignment. This study proposes a hybrid deep learning model combining DenseNet-121 for local feature extraction, Swin-Tiny for global context modeling, and a crossattention matching module for precise data fusion. The methodology involves preprocessing (normalization, resizing to 224x224, and augmentation), feature extraction, hierarchical refinement, and similaritybased alignment. Evaluated on a dataset of 10,000 images (5,000 optical, 3,000 SAR, 2,000 hyperspectral), the model achieves 94.6% accuracy, 88.7% SSIM, and 15.4% RMSE, outperforming DenseNet-only (89.5% accuracy, 82.3% SSIM, 19.8% RMSE) and Swin Transformer-only (91.0% accuracy, 85.1% SSIM, 17.2% RMSE) baselines. It also surpasses state-of-the-art methods like SwinV2DNet (92.3%) and STransFuse (90.8%) by 2.3-3.8% in accuracy. With an inference time of 0.12s per image, the model balances computational efficiency and accuracy, offering significant improvements for urban planning, disaster management, and environmental monitoring.