TerraDA: A Domain-Adaptive Framework With Dynamic Multimodal Remote Sensing Data Fusion and Consistency Learning for Cross-City Land Cover Classification

Land cover classification plays a critical role in urban planning, resource management, and disaster assessment. Recent advances in deep learning have substantially enhanced remote sensing-based land cover classification, particularly for urban environments. However, cross-city generalization of existing models is constrained by distribution shifts in remote sensing data, multimodal heterogeneity, and scarce semantic annotations in target cities. To address these challenges, we propose TerraDA, a novel framework for cross-city multimodal remote sensing land cover classification. TerraDA systematically integrates three core modules: A Fine-Grained Attention-Modulated Encoder (FAME), a Multi-Scale Dynamic Gated Feature Fusion Module (MS-DGF), and a Dynamic Consistency-Adversarial Adaptation Module (DCAA). Specifically, FAME enhances the modality-specific feature representations of hyperspectral imaging, multispectral imaging, and synthetic aperture radar data through fine-grained spatial attention; MS-DGF adaptively fuses multimodal features via a dynamic gating mechanism while capturing multiscale spatial context; DCAA introduces dynamic weighting and a joint consistency-adversarial loss to robustly align cross-domain features. Experiments on the C2Seg-AB and C2Seg-BW datasets demonstrate TerraDA’s superior performance, achieving mIoU/F1/OA/Kappa scores of 26.36/37.02/59.71/50.14% and 11.66/17.64/38.45/29.06%, respectively, surpassing state-of-the-art methods, including SegNet, PSPNet, DeepLabV3+, FastFCN, SegFormer, and HighDAN, in cross-city land cover classification.