Leveraging Synthetic Data and Machine Learning for Cloud Optical Thickness Estimation

Cloud formations often obstruct the effectiveness of optical satellite monitoring, imposing limitations on Earth observation (EO) tasks such as land cover mapping, ocean color analysis, and cropland monitoring. While machine learning (ML) methods have improved EO tasks, challenges persist, primarily the dependence on annotated data for ML training, especially in EO contexts like cloud optical thickness (COT) estimation. To address the scarcity of COT data, we propose a synthetic dataset simulating top-of-atmosphere radiances for 12 spectral bands of the MSI sensor on Sentinel-2 platforms, and encompassing various cloud types, COTs, and environmental conditions. Extensive experimentation on training ML models to predict COT from spectral band reflectivities demonstrates the utility of the proposed dataset. Generalization to cloud cover mapping on real data is verified on two satellite image datasets. The data, code and models will be made available.