Daily Cloud-Free NDSI Reconstruction Across the Northern Hemisphere Driven by Nonlinear Meteorological Forcing

Persistent cloud contamination in MODIS normalized difference snow index (NDSI) products severely limits reliable snow cover monitoring across the Northern Hemisphere, making effective gap-filling crucial. However, existing approaches often oversimplify snow temporal dynamics and fail to capture the cumulative nature of snow–climate interactions. To address these limitations, we propose PredFormer, a novel sequence-based framework that extends the self-attention mechanism to the temporal dimension, thereby explicitly modeling the nonlinear temporal evolution of NDSI coupled with cumulative meteorological effects. The framework further incorporates vegetation growth, topographic conditions, and cloud mask information to account for complex environmental dependencies. Validation results demonstrate that PredFormer achieves superior reconstruction performance across the Northern Hemisphere, with MAE and RMSE values of 1.110 and 2.946, respectively. Notably, the distinct mechanism of incorporating nonlinear cumulative meteorological effects reduces RMSE by 11.57%, and comparative analyses against baseline models (e.g., LSTM and standard Transformers) reveal performance gains exceeding 19%. The framework also demonstrates substantial improvements in high-elevation and forested regions—including the Hindukush–Himalaya and the West Coast. Leveraging this framework, we generate the first hemispheric-scale, daily cloud-free NDSI dataset (5-km resolution, 2003–2023). This work not only advances the methodological handling of nonlinear snow dynamics but also delivers a foundational dataset for hydrological modeling and climate change assessment.