Automatic water detection from multidimensional hierarchical clustering for Sentinel-2 images and a comparison with Level 2A processors

Continuous monitoring of water surfaces is essential for resource management. This study presents a nonparametric unsupervised automatic algorithm the identification inland pixels from multispectral satellite data using multidimensional clustering and high-performance subsampling approach large scenes. Clustering analysis technique that used to identify similar samples in space. The spectral information derived indices were characterize each scene pixel individually. A machine learning with random generalization through Naïve Bayes classifier was also proposed make application complex algorithms scenes feasible. Accuracy evaluated an independent dataset provides bodies 15 Sentinel-2 images over France acquired different seasons covers range colour types. validation surface more than 1200 km2 (approximately 12 million pixels) including 80,000 outlined semiautomatic active method, which manually revised. classification results compared three major Level 2A processors (MAJA, Sen2Cor FMask) two most common thresholding techniques: Otsu Canny-edge. An input mask remove coastal waters, clouds, shadows snow pixels. Water identified automatically process without need ancillary or pretrained data. Combinations up (Modified Normalized Difference Index-MNDWI, Index-NDWI Multiband Index-MBWI) reflectance bands (B8 B12) tested algorithm, best combination NDWI-B12. Of all methods, our method achieved highest mean kappa score, 0.874, across scenes, per-scene ranging 0.608 0.980, lowest standard deviation 0.091. Standard Otsu's had worst performance due lack bimodal histogram, Canny-edge variation overall 0.718 when MNDWI. For masks provided by generic processors, FMask outperformed MAJA obtained 0.764. In-depth shows quick drop methods identifying area below 0.5 ha, but second 34% this size class.