KD-Assim: Incremental Kalman-Delta Assimilation of Multi-Source Earth Observations under Cloud Contamination and Budget Constraints
Year: 2026
Type: preprint
Abstract: Persistent cloud and cloud-shadow coverage breaks optical satellite time series exactly when environmental monitoring needs them most: drought assessment requires dekadal vegetation trajectories, and flood mapping requires water extent within days, yet a single overcast acquisition can remove the decisive date from a season. Multi-modal reconstruction with Sentinel-1 synthetic aperture radar (SAR) is the strongest available remedy, but current models fuse modalities through attention weights, which are relative, physically uninterpretable, and unable to express that both sensors are unreliable; they also encode the whole sequence at once, so compute grows with the number of acquisitions and cannot run on board. We recast cloud-affected reconstruction as incremental data assimilation over Earth observations and contribute three components. Contribution-aware covariance modulation (CACoM) reparameterises modality contribution as a per-pixel, per-channel observation precision, so fusion becomes a precision-weighted Bayesian update and missing observations receive zero precision instead of being extrapolated. The Kalman-delta gain network (KDGNet) replaces sequence encoding with a predict–update recursion whose transition is a learned delta and whose gain is a low-rank factor computed from the innovation, keeping compute linear in the number of acquisitions. Budget-aware uncertainty gating (BAUG) spends refinement steps only where the posterior variance is high, turning a fixedcost black box into an anytime estimator with an explicit usable-pixel mask. On SEN12MS-CR- TS, KD-Assim reaches 28.52 dB PSNR, 0.880 SSIM and 9.86◦ SAM, improving 0.68 dB over the strongest single-model baseline while using 52% of its FLOPs; it transfers to real clouds in AllClear at 29.61 dB, and its predicted spread is better calibrated than a five-member deep ensemble at one fifth of the inference cost.
Language: English
Subfield: Media Technology
Field: Engineering
Domain: Physical Sciences
SDG: Climate action
Open Access status: green
OpenAlex ID: W7225991467
