Reconstructing cloud-free Sentinel-2 time series under complex degradations with state-driven spatio-temporal modeling

Satellite Image Time Series (SITS) are indispensable for monitoring complex, long-term Earth surface dynamics. However, the temporal continuity of high temporal frequency optical observations is frequently compromised by severe degradations such as clouds, blur, and sensor noise. Deep learning has driven progress in time-series reconstruction through powerful spatio-temporal modeling. Nevertheless, existing models typically treat this as an isolated cloud-removal task, and perform unconstrained spatio-temporal information aggregation throughout the sequences. These approaches insufficiently model continuous temporal correlation and ignore whether observations are clear or degraded, allowing temporally uncorrelated or degraded features to contaminate the feature aggregation. In this paper, we propose a State-driven Spatio-Temporal Reconstruction approach through Directed Information Routing (STRIDE), designed for the robust reconstruction of long-term SITS under complex degradations. First, the state-aware directed information routing (SDIR) mechanism converts degradation masks into state-aware routing biases. It is introduced to promote information flow from clean observations to degraded ones, and inhibit degraded features from contaminating clean observations. Second, the manifold-mapped temporal evolution (MMTE) mechanism evaluates temporal correlation by balancing linear succession and seasonal periodicity, guiding the network to prioritize temporally relevant features during temporal feature aggregation. Driven by these two mechanisms, the asymmetric spatio-temporal transformer (ASTT) modulates self-attention based on the level of degradation and temporal correlation, encouraging directional feature aggregation. Finally, STRIDE employs an auxiliary degradation estimation module for non-cloud degradations and optionally integrates complementary sensor data, such as synthetic aperture radar (SAR) data. Extensive evaluations demonstrate that our approach achieves robust SITS reconstruction in global scenarios.