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

Adrián Cal 

Cloud shadows constitute a significant source of uncertainty in optical remote sensing, thereby impairing the accuracy of spectral indices, land cover classifications, and time-series analyses derived from Sentinel-2 imagery. Although the Sentinel-2 Level-2A Scene Classification Layer (SCL) offers information on cloud shadows, its efficacy is frequently hindered by high omission rates and inconsistent detection. This study presents a novel spectral index, namely, the Cloud Shadow Detection Index (CLOSDI), which is specifically designed to enhance cloud shadow detection within agricultural and grassland landscapes.

CLOSDI is derived from the differential sensitivity of the Normalized Difference Vegetation Index (NDVI) and the two-band Enhanced Vegetation Index (EVI2) to shadow conditions and can be computed directly from RED and NIR reflectance using a closed-form formulation. The index was evaluated using 1,231 high-quality image patches from the CloudSEN12 dataset, which are restricted to temperate and tropical grassland biomes. An optimal cutoff threshold of 34.0 was identified through an 80/20 training–testing split by maximizing the median intersection over union (IoU).

On the test set, CLOSDI substantially outperformed the Sentinel-2 SCL-based shadow mask. The median performance metrics were: recall of 76.8 versus 19.6, F1-score of 62.1 versus 19.6, IoU of 45.0 versus 16.7, and balanced overall accuracy of 80.1 versus 57.6 for CLOSDI and SCL, respectively. A Wilcoxon signed-rank test confirmed that the improvement in IoU was statistically significant (p = 1.402 × 10⁻¹⁹). Despite using only two spectral bands, CLOSDI achieved a performance comparable to state-of-the-art cloud shadow detection algorithms, while requiring substantially lower computational complexity and eliminating the need for model training.

These findings illustrate that the CLOSDI offers a straightforward, physically interpretable, and computationally efficient approach for cloud shadow detection within vegetation-dominated landscapes. Its ease of implementation and robust performance render it particularly appropriate for large-scale agricultural monitoring, vegetation time-series analysis, and operational remote sensing workflows.