VegFloodv2.0: dataset for flood detection beneath the vegetation cover

VegFlood is a reference dataset designed for training, validation, and testing of deep learning models for semantic segmentation of flooded areas — including regions covered by vegetation. The dataset was built using publicly available Sentinel-1 synthetic aperture radar (SAR) imagery, high-resolution digital terrain models (DTMs), and hydrological observations from river gauging stations. It includes 1,707 image–mask pairs collected from 39 gauging stations, geographically limited to the territory of Poland. 
The SAR images are derived from Sentinel-1 Ground Range Detected (GRD) products processed through the Alaska Satellite Facility's HyP3 service. Each image is stored in GeoTIFF format with three bands: VV backscatter, VH backscatter and VV/VH ratio. All bands are stored as float32 arrays. To remove noise, the VV backscatter values were clipped to the range (−23, 0 dB) and VH backscatter values were clipped to the range (−28, −5 dB). Values outside these ranges were excluded. All bands were then normalized to the [0, 1] range using min-max normalization.
Multi-class label masks were generated by combining information from a DTM-based water mask (obtained by thresholding the DTM on the observed water level at each gauging station on the acquisition date) and a water mask derived from preprocessed SAR images. These annotated data are stored in GeoTIFF format, encoded as uint8 arrays: 0 – non-flooded class, 1 - open-water class, 2 - flooded vegetation class.
To ensure spatial independence between training and evaluation data, the dataset was split by gauging station: training set: 28 stations, validation set: 5 stations, testing set: 6 stations.