S1S2-Water: A global dataset for semantic segmentation of water bodies
from Sentinel-1 and Sentinel-2 satellite images
Abstract
This study introduces the S1S2-Water dataset – a global reference
dataset for training, validation and testing of convolutional neural
networks for semantic segmentation of surface water bodies in publicly
available Sentinel-1 and Sentinel-2 satellite images. The dataset
consists of 65 triplets of Sentinel-1 and Sentinel-2 images with quality
checked binary water mask. Samples are drawn globally on the basis of
the Sentinel-2 tile-grid (100 x 100 km) under consideration of
predominant landcover and availability of water bodies. Each sample is
complemented with metadata and Digital Elevation Model (DEM) raster from
the Copernicus DEM. On the basis of this dataset we carry out
performance evaluation of convolutional neural network architectures to
segment surface water bodies from Sentinel-1 and Sentinel-2 images. We
specifically evaluate the influence of image bands, elevation features
and data augmentation on the segmentation performance and identify
best-performing baseline-models. The model for Sentinel-1 achieves an
Intersection Over Union of 0.85, Precision of 0.93 and Recall of 0.90 on
the test data. For Sentinel-2 the best model produces an Intersection
Over Union of 0.96, Precision of 0.99 and Recall of 0.97 respectively.
We also evaluate the performance impact when a model is trained on
permanent water data and applied to independent test scenes of floods.