Spectral Characteristics of the Dynamic World Land Cover Classification

The Dynamic World product is a discrete land cover classification of Sentinel 2 reflectance imagery that global in extent, retrospective to 2015, and updated continuously near real time. classifier trained on stratified random sample 20,000 hand-labeled 5 × km tiles spanning 14 biomes globally. Since the training data are based visual interpretation image composites by both expert non-expert annotators, without explicit spectral properties specified class definitions, characteristics classes not obvious. objective this study quantify physical distinctions among characterizing range present within each over variety landscapes. This achieved comparing eight-class probability feature space (excluding snow) maximum assignment (label) distributions continuous fraction estimates derived from globally standardized mixture model. Standardized substrate, vegetation, dark (SVD) endmembers used unmix nine diversity hotspots for comparison between SVD continua assignments. variance partition spaces indicates eight these effectively five-dimensional 95% variance. Class all show tetrahedral form with multiple classes. Comparison assignments (labels) reveal clear distinction (1) physically spectrally heterogeneous characterized gradations vegetation density, substrate albedo, structural shadow fractions, (2) more homogeneous closed canopy (forest) or negligible (e.g., desert, water). Due ubiquity worldwide, adds considerable value labels offering users opportunity depict inherently gradational landscapes otherwise generally offered other thematic classifications.