An Automated Cloud Detection Method for Sentinel-2 Imageries

With the tremendous boost of Remote Sensing technology for different Earth observation activities, the need for fast and automated extraction of contamination caused by clouds is of considerable importance to exploit the potential of freely available satellite imagery like Sentinel-2. Several cloud detection (CD) methods have extensively been used with remote sensing data, including threshold-(Fmask and Sen2Cor) and Machine Learning (ML) based methods. ML cloud detection resolved the problem of poor universality faced by threshold-based CD approaches. In spite of the advantages and popularity of extensively used ML methods (Support Vector Machine (SVM) and Random Forest classifier), these require prior feature extraction using methods to include spatial features with spectral features during classification, requiring a higher computational cost. The use of XGBoost, another ML-based classifier, which is a relatively faster and highly accurate alternative to SVM and RF classifiers, has been explored for cloud detection. Similar to other classifiers XGBoost also requires additional spatial and spectral features to achieve state-of-the-art results. Keeping in view of the requirement of spatial information, this study proposed a novel patch-based cloud detection method based on ML algorithms for Sentinel-2 imagery using the benchmarked Baetens-Hagolle dataset (CESBIO/CNES). All three machine learning classifiers, i.e., XGBoost, Random Forest, and SVM, were used for classification to estimate the proposal’s effectiveness. A comparison of results suggests an improved performance by the XGBoost-based cloud detection method in terms of accuracy and other parameters. XGBoost achieved accuracy, and F1 score values of 90.26% & 78.77%, respectively, and it required under 4 seconds entirely to generate the cloud mask per imagery, which surpasses the speed of Random Forest and SVM classifiers by a factor of 6 & 250, respectively.