MAFFUNet: A Multimodal Attention and Feature Fusion Framework for Remote Sensing Image Integration

Multimodal image fusion in remote sensing, integrating data from sources like SAR, optical, multispectral, and hyperspectral sensors, is critical for applications such as land cover classification and disaster monitoring. However, challenges like modality heterogeneity, cloud occlusion, and cross-domain variability demand advanced fusion algorithms capable of capturing nonlinear correlations and ensuring robust generalization.We propose MAFFUNet, a lightweight U-Net-based framework with a five-layer encoder-decoder architecture, incorporating depthwise separable convolutions for efficiency and temporal data support via flattening. MAFFUNet integrates three novel modules: Adversarial Feature Fusion (AFF) for maximizing mutual information, Hybrid Attention and Multiscale Fusion (HAMF) for optimizing spectral and spatial features, and Cross-Domain Regularization (CDR) for enhancing cross-domain adaptability.Experiments on the Chikusei, PaviaC, and PaviaU datasets at 4×, 8×, and 16× scaling factors demonstrate that MAFFUNet outperforms traditional methods and deep learning approaches in metrics like RMSE, PSNR, ERGAS, and SAM, with superior spectral fidelity and spatial detail preservation. Its low computational complexity suits resource-constrained environments.MAFFUNet provides an efficient and robust solution for multimodal remote sensing image fusion.