Densification and forecasting of Sentinel-2 time series from multimodal SAR and Optical satellite data using deep generative models

Optical satellite imagery is essential for agriculture, climate monitoring, and land surface analysis, but clouds and swath edges create irregular sampling that breaks temporal continuity. Existing methods only reconstruct missing observations within the observed time window—they cannot predict future imagery. A new paper from Defonte, Derksen, Constantin, and Nespoulous (arXiv:2605.04239) tackles both problems with a deep generative model that leverages multimodal data: Sentinel-2 optical and Sentinel-1 SAR (synthetic aperture radar, which penetrates clouds). The framework probabilistically generates optical images at any requested timestamp, either filling historical gaps or forecasting future scenes. This goes beyond simple interpolation by learning the underlying temporal dynamics from the complementary SAR signal, which is unaffected by weather.

A key innovation is the focus on uncertainty quantification. Most prior generative approaches for satellite imagery produce deterministic outputs, but this model outputs a distribution, letting users know how confident the system is in each generated pixel. Experiments show strong performance on sparse and temporally misaligned time series, demonstrating both densification (filling missing dates) and forecasting (predicting future dates) with realistic image quality. The work opens the door to near-real-time continuous monitoring even in persistently cloudy regions—critical for crop yield prediction, deforestation tracking, and disaster response. The code and data are expected to be released pending publication.