Cloud-Free Imaging Probability Forecasting for Optical Earth-Observation Tasking over Yerevan, Armenia

Isabella Avasapian

Optical Earth-observation satellites such as Sentinel-2 cannot see through cloud, so a tasking attempt over a cloudy target wastes a limited imaging window along with onboard power and downlink bandwidth. We study whether next-day cloud conditions over Yerevan, Armenia can be forecast accurately enough to automate the binary Go/No-Go tasking decision. Daily cloud-free fractions are derived from the Sentinel-2 Level-2A scene classification for 933 days spanning 2020–2024 and paired with ten ERA5 atmospheric reanalysis variables; four days of atmospheric context (today and three lags) form a 40-feature predictor for the following day’s cloud-free fraction. Using strictly time-ordered cross-validation, a gradient-boosted tree model (XGBoost) attains a mean absolute error of 0.165 on the cloud-free fraction and a random forest attains 0.160, improving on a persistence baseline (0.169) by 2.4–5.5%. Because cloudiness over the orographically stable Ararat Valley is strongly autocorrelated, persistence is a demanding benchmark, and the learned gain reflects a genuine atmospheric signal rather than trend-following. An operational variant generalizes the approach to multiple Armenian regions, is calibrated as a probabilistic Go/No-Go classifier (cross-validated AUC 0.77, Brier score 0.17 over 5,736 samples), and is driven at inference time by the cloud-free, no-key Open-Meteo forecast API to sidestep the production latency of ERA5. We release the system as an openly accessible high-frequency cloud-imaging forecast for the South Caucasus, a tool that to our knowledge does not otherwise exist for the region.