Multi-source remote sensing data fusion based on attention-enhanced deep learning for intelligent land cover classification

Abstract

The rapid growth of multi-source Earth observation data has introduced significant challenges for land cover classification. These challenges arise from two factors: cross-modal distribution gaps and spatial misalignment between optical imagery and synthetic aperture radar (SAR) data. This paper proposes AE-Net, an Attention-Enhanced Deep Learning Network for optical-SAR fusion. AE-Net introduces three key innovations. First, the Cross-Modal Feature Alignment (CMFA) module bridges distribution gaps via adaptive instance normalization and geometric offset correction. Second, the Decoupled Channel-Spatial Attention Module (DC-SAM) employs modality-specific excitation pathways to preserve heterogeneous statistical properties. Third, the Uncertainty-Guided Fusion Pyramid (UGFP) adaptively weights modalities by prediction confidence through Monte Carlo dropout. Experiments on the MSRSD dataset (256 × 256 patches) demonstrate that AENet achieves 96.8% overall accuracy and a Kappa coefficient of 0.963, significantly outperforming nine state-of-the-art methods including UDFNet (95.1%), PICNet (94.7%), and MultiModNet (94.3%) (paired t-test, p < 0.01). Ablation studies confirm that CMFA provides the largest single gain of 2.1% in overall accuracy over early fusion.