Relaxing the clear-sky assumption: Cloud-tolerant spatiotemporal fusion via mask-guided feature modulation and temporal-memory collaboration

The precise characterization of dynamic Earth surface processes relies on time-series remote sensing imagery with both high spatial fidelity and frequent temporal coverage. While spatiotemporal fusion (STF) bridges this resolution trade-off by integrating complementary observations, its practical application remains constrained by the idealized ”clear-sky assumption”. This requirement often favors the selection of temporally distant cloud-free acquisitions over temporally closer but cloud-contaminated observations, potentially reducing the relevance of reference information. To address this limitation, we propose CloudSTF, a robust framework designed to achieve high-fidelity reconstruction by effectively exploiting partially occluded observations. Our approach employs a Mask-guided Multi-scale Swin Transformer (M 2 ST) encoder to capture cloud spatial patterns and suppress noise propagation during feature extraction. This is coupled with a Cross-temporal Memory-guided Fusion (CTMF) module that adaptively integrates temporal trends with textural details retrieved from a spatio-temporal memory bank. To validate this paradigm, we introduce the Global Cloud-shrouded Regions (GCR-STF) benchmark, a geographically distributed dataset comprising 24 representative agricultural and urban sites across six continents. Extensive experiments demonstrate that CloudSTF consistently outperforms state-of-the-art methods. Additional assessments across varying cloud densities and diverse landscapes further demonstrate the robustness and applicability of CloudSTF under realistic cloud-contaminated conditions. These results suggest that CloudSTF provides a reliable solution for continuous Earth monitoring in realistic, cloud-prone environments. Source code and the GCR-STF benchmark are available at https://github.com/Sichen-Lu/CloudSTF .