RTFormer: radiative transfer model-coupled transformer for cloud removal in optical remote sensing imagery

Clouds cover approximately 70% of the Earth’s surface, severely degrading optical satellite imagery and limiting its use in critical applications. Although recent transformer-based cloud removal methods have shown promise, they face two key challenges: (1) empirical models are limited in representing the complex radiative transfer properties of clouds, leading to physical inconsistencies; and (2) the high computational demands of self-attention typically constrain receptive fields, hindering the exploitation of global context for reconstructing cloud-contaminated regions. To address these issues, we propose RTFormer, a novel Radiative Transfer Model-coupled Transformer framework that integrates physical modeling into deep learning for efficient and accurate cloud removal. The core contribution lies in the incorporation of radiative transfer principles into the transformer architecture, enabling global context perception without additional computational overhead. To bridge the gap between simulated and real-world scenarios, we furhter developed a radiative transfer model-based dataset for model training. RTFormer extends the perception field beyond local constraints and incorporates a residual learning mechanism to preserve intact information in clear regions while accurately reconstructing cloud-affected areas. Experimental results demonstrate that RTFormer outperforms state-of-the-art methods in accuracy. Moreover, it substantially benefits downstream applications; for instance, when applied as a preprocessing step for land-cover classification, the overall accuracy improved by 4.38% compared with competing methods. These findings highlight the value of integrating physical models with transformer-based deep learning for cloud removal and demonstrate its potential for advancing remote sensing applications.