Global Sentinel-2 Snow Semantic Segmentation Dataset with Dynamic World Labels

The dataset is designed for snow semantic segmentation using Sentinel-2 satellite imagery. It contains multispectral images with 13 spectral bands (Top-of-Atmosphere reflectance) along with corresponding binary snow masks derived from Dynamic World land cover classifications. The dataset was generated using Google Earth Engine from the “COPERNICUS/S2_HARMONIZED” and “GOOGLE/DYNAMICWORLD/V1” collections. A total of 8300 random locations across the Earth's surface were sampled. For each location, a bounding box of varying size and a random temporal interval (from 2023 onward) were defined. Within this interval, an available Sentinel-2 image with up to 5% cloud coverage was selected. For each sample, a multispectral Sentinel-2 image and the corresponding Dynamic World segmentation map were extracted for the same spatial extent. The Dynamic World labels were then converted into binary snow masks, where snow is labeled as 1 and all other classes as 0. The dataset is intended for training and evaluation of machine learning models for snow detection and semantic segmentation tasks. The dataset is organized as follows: Raw georeferenced Sentinel-2 images are stored in the /raw directory, corresponding binary snow mask images are stored in the /snow_labels directory, and a JSON file containing metadata and geographic locations of all samples is included.