Iraq Lake image dataset for semantic segmentation

This dataset contains satellite imagery of major lakes in Iraq, including Lake Razzaza, Lake Habbaniyah, Lake Tharthar, Lake Dukan, and Lake Darbandikhan. The dataset is designed for training and evaluating fully convolutional neural networks for semantic segmentation of surface water from Sentinel-2 imagery. Sentinel-2 Level-2A (surface reflectance) data were accessed and processed using Google Earth Engine. For each lake area of interest and target year, seasonal median composites were generated to reduce the effects of shoreline blurring, cloud contamination, and seasonal water–land mixing. Four 10 m spatial resolution bands were exported as GeoTIFF images: B4 (Red), B3 (Green), B2 (Blue), and B8 (Near-Infrared). All bands were scaled to 8-bit values (0–255) for compatibility with convolutional neural network inputs. Ground-truth water masks were generated using the JRC Global Surface Water YearlyHistory product. Pixels classified as seasonal water (class 2) and permanent water (class 3) were treated as the positive class. In some regions, masks were further refined using a Normalised Difference Water Index (NDWI) threshold to reduce false positives over land. The resulting image–mask pairs were tiled into 512×512 pixel patches and organised into training, validation, and test subsets for supervised semantic segmentation. Binary masks are provided as PNG files, where water pixels are labelled as 1 (white) and background pixels as 0 (black). Dataset structure: Training set: 857 image-mask pairs Validation set: 340 image-mask pairs Test set: 63 image-mask pairs All image tiles are provided as GeoTIFF files with a spatial resolution of 10 m per pixel. Data sources: Sentinel-2 Level-2A imagery (Copernicus/ESA) JRC Global Surface Water Dataset This dataset is intended for research and educational use in remote sensing, hydrology, and machine-learning-based water body mapping.