Beyond the SCL: threshold optimization of Cloud Probability and Cloud Score+ for cloud and shadow masking in Sentinel-2 imagery using CloudSEN12

Adrián Cal , Guadalupe Tiscornia

Cloud and cloud shadow contamination remains a major source of uncertainty in Sentinel-2 imagery. The native Scene Classification Layer (SCL) frequently exhibits high omission errors and offers no threshold flexibility, whereas Cloud Probability (CP) and Cloud Score+ (CS+) provide continuous, user-adjustable scores. However, no consensus exists on their optimal thresholds, and their capacity for joint cloud-and-shadow masking remains largely unquantified. This study identifies optimal masking thresholds (OMTs) for CP and the cs and cs_cdf bands of CS+, for both cloud-only (BCM) and joint cloud-and-shadow (BCSM) masking, and compares their performance against SCL. Using 3,994 high-quality CloudSEN12 patches split by region of interest and stratified by land cover, OMTs were determined on a training subset by maximizing median intersection over union (IoU) and evaluated on 804 independent test patches. The CP-BCM threshold (37.7) closely matched the algorithm’s recommended default, providing independent empirical validation. Across all six product-mask type combinations, cs_cdf-BCSM achieved the highest performance (OMT = 76.4, equivalent to 0.764 in the native 0–1 scale; F1-Score = 90.0%; IoU = 81.9%; balanced overall accuracy = 87.8%), confirmed as statistically significant by Friedman and Nemenyi tests. CS+ products improved substantially when transitioning from cloud-only to joint masking, whereas CP degraded and SCL recorded the lowest performance in both scenarios. These results proved robust across six independent data partitions and generalized consistently across ten land-cover classes. These findings provide reproducible threshold recommendations for cloud and cloud shadow masking in Sentinel-2 imagery, establishing cs_cdf as the most effective product for joint masking applications.