AdaptCloud: Information-Driven Adaptive Multimodal Fusion for Cloud-Robust Land Cover Classification

Land cover classification (LCC) is essential for resource management and socioeconomic analysis but is often hindered by clouds that obscure spectra and disrupt spatial–temporal consistency. Current optical and SAR fusion methods aimed at mitigating these impacts lack cloud conditioned reliability modeling, which restricts their generalization ability. To address this issue, we propose AdaptCloud, an information-driven adaptive multimodal fusion framework designed for cloud robust LCC. AdaptCloud mitigates cloud induced information degradation by constructing an information exploitation chain that treats cloud cover as a limited information condition, progressively enhancing class discriminability, observational completeness, and overall reliability. Specifically, for enhanced category discrimination, AdaptCloud leverages prior information from foundational model to strengthen feature extraction capabilities under cloud occlusion, while category adaptive fine-tuning optimizes feature extraction for specific category. By leveraging occlusion stable low frequency optical structures and high frequency SAR texture priors, AdaptCloud performs bidirectional injection and pixel level multiscale fusion to improve information integrity. Furthermore, to enhance key information reliability, AdaptCloud aligns modalities via cross modality attention and integrates sparse filtering to preserve stable features and suppress degradation, enabling robust multimodal fusion under diverse cloud conditions. Systematic evaluations on four cloud occluded datasets, DDHRNet Korea, DDHRNet Xi’an, PIE-RGB-SAR, and CSU-OPT-SAR, show that AdaptCloud outperforms state-of-the-art baselines for land-cover mapping under cloud cover, achieving mIoU gains of 0.71% to 19.64% and demonstrating robust performance in cloud prone regions. Additionally, the CSU-OPT-SAR dataset is established to overcome the lack of multi-seasonal data, demonstrating AdaptCloud’s strong generalization for cross-regional and cross-temporal LCC.