Sentinel-2 Surface Water Segmentation Dataset for the Southern Iraqi Marshes (2021)

Dataset Description This dataset contains Sentinel-2 satellite imagery of the Southern Iraqi Marshes, prepared for training and evaluating fully convolutional neural networks (FCNs) for binary semantic segmentation of surface water. The dataset is intended for research and educational use in remote sensing, hydrology, and machine-learning-based water body mapping, particularly for benchmarking deep learning models on medium-resolution multispectral imagery. Data Sources Sentinel-2 Level-2A surface reflectance imagery (Copernicus / ESA), accessed and exported using Google Earth Engine JRC Global Surface Water (GSW) Yearly History dataset for water extent labels Image Preprocessing Sentinel-2 imagery was exported as four 10 m spatial resolution bands: B4 – Red B3 – Green B2 – Blue B8 – Near-Infrared The bands were exported as 8-bit GeoTIFF files and subsequently processed offline. All image–mask pairs were spatially aligned to ensure identical coordinate reference systems, transforms, and pixel grids. Where required, masks were resampled to the image grid using nearest-neighbour reprojection to preserve discrete class labels. Each image–mask pair was tiled into non-overlapping 512 × 512 pixel patches. Image tiles were then min–max normalised on a per-tile basis and stored as floating-point arrays in the range [0, 1]. Final image tiles have shape (512, 512, 4) and dtype float32. Ground-Truth Water Masks Water masks were derived from the JRC Global Surface Water Yearly History product. Exported water layers were binarised such that: Water pixels → 1 Background (non-water) pixels → 0 Masks were saved as NumPy arrays with shape (512, 512, 1) and dtype uint8. This release uses 2021 data only, as configured in the preprocessing pipeline. Dataset Splitting and Balancing To mitigate severe class imbalance commonly present in water segmentation tasks, tiles were categorised based on water coverage fraction: Empty: 0% water Mixed: >0% and <30% water Water-heavy: ≥30% water A balanced subset of tiles was sampled into training, validation, and test splits using fixed proportions of empty, mixed, and water-heavy tiles. Splits are non-overlapping at the tile level, ensuring no spatial leakage between subsets. Validation and test sets were additionally constrained to include a minimum number of water-heavy tiles where available. Final split sizes: Training set: 250 image–mask pairs Validation set: 100 image–mask pairs Test set: 20 image–mask pairs Intended Use This dataset is designed for: Training and evaluation of semantic segmentation models (e.g. U-Net, SegNet, Attention-U-Net) Research on water body detection, wetland monitoring, and environmental change analysis Benchmarking machine learning pipelines on multispectral satellite imagery Licence and Usage This dataset is released for non-commercial research and educational purposes. Users should cite the original data sources (Copernicus Sentinel-2 and JRC Global Surface Water) when using this dataset in academic work.