Landsat-based Water–Ice Reference Dataset for the Vistula River

This dataset contains training and test data used for a Random Forest–based classification of Landsat satellite imagery for river ice detection. It is derived from multispectral Landsat surface reflectance data and includes a set of spectral bands and features designed to improve the separability between ice and open water.

The dataset consists of pixel-level observations extracted from Landsat scenes and includes predictor variables such as Green, Near-Infrared (Nir), NDSSI, band ratios (GB, RG), HSV color space components (hue, saturation, value), and the WICI index. Each record is associated with a Landsat scene identifier (scene_id), a binary target variable indicating ice or water presence (class), and a train_test flag used for the initial split into training and test subsets.

The data were prepared to support supervised machine learning experiments, specifically Random Forest classification, and to evaluate the separability of river ice and water based on spectral and derived features.

The creation of the dataset was funded in whole or in part by the National Science Centre, Poland (grant no. 2025/57/N/ST10/00545) and the Polish National Agency for Academic Exchange (NAWA Preludium Bis 2 programme).