Physics-informed machine learning for cloud detection

Accurate cloud detection is a critical preprocessing step for analyzing optical remote sensing data, such as from Landsat and Sentinel-2. In this study, we developed Fmask 5 (Version 5 of Function of mask) for cloud detection in Landsats 4–9 (Landsat 4, 5, 7, 8, and 9) and Sentinel-2 (Sentinel-2 A, B, and C) imagery, based on a novel physics-informed machine learning (PIML) framework. This framework synergistically integrates physical rules with machine learning (ML) models, overcoming their respective limitations. Specifically, a base ML model first generates a preliminary cloud mask to adapt to the application of physical rules, determining the optimal combination of spectral tests and thresholds for that image. The resulting physical-rule-adapted mask serves as additional “localized” training data to fine-tune the ML model. For a global, independent validation dataset combining simple random and disagreement-stratified samples from over 600 images per sensor, Fmask 5 achieved overall accuracies of 93.46% for Landsats 8–9, 92.50% for Landsats 4–7, and 95.38% for Sentinel-2, representing improvements of 1.6–6.0 percentage points over Fmask 4. The Fmask 5 algorithm is CPU-efficient, making it suitable for large-scale operational processing.