CLOSDI: A Novel Spectral Index for Cloud Shadow Detection in Sentinel-2 Imagery using NDVI and EVI2

Adrián Cal 

The presence of clouds and their shadows represents one of the main limitations for the spectral analysis of Sentinel-2 imagery. Although the Scene Classification Layer (SCL), generated by the Sen2Cor algorithm, includes specific classes for cloud shadows and dark area pixels, several studies have revealed limitations in its detection capabilities. This work proposes a new spectral index—the Cloud Shadow Detection Index (CLOSDI)—based on the relationship between NDVI and EVI2, aimed at improving the identification of pixels affected by cloud shadows.
Using the CloudSEN12 dataset as a reference, binary shadow masks were generated from CLOSDI and compared with those provided by the SCL (referred to as S2A-BSM). The optimal CLOSDI Cutoff Threshold (CCT = 35) was determined through an automated process over 3,161 training patches and subsequently applied to an independent test set of 300 patches. The masks were evaluated using standard segmentation metrics: Precision, Recall, F1-Score, IoU, and Balanced Overall Accuracy (BOA).
The masks generated with CLOSDI (CLOSDI-BSM) significantly outperformed those from S2A-BSM across all metrics, achieving a BOA of 76.6 compared to 49.7 for S2A-BSM. When compared to other methods reported in CloudSEN12, CLOSDI-BSM ranked third overall, only behind the Human Level classification and the deep learning-based UNetMobV2 model, and above well-established algorithms such as Fmask, KappaMask, and Sen2Cor.
Unlike these methods, CLOSDI requires no training or GPU usage, and offers a computationally lightweight implementation, making it an effective and accessible tool for cloud shadow detection in Sentinel-2 products, particularly in operational contexts or resource-constrained environments.