High-Resolution Remote Sensing Dataset for Levee-Like Feature Segmentation in Arid Regions

The dataset includes 459 single-phase reference images at approximately 0.5-meter spatial resolution with corresponding pixel-level annotations, as well as multi-temporal images of the same areas captured at different times, totaling 1550 images. All annotations were manually drawn using the LabelMe tool in polygon mode and converted to YOLO-format label files (.txt).

A spatial location-based data splitting strategy was adopted to prevent data leakage: the training set (~80%), validation set (~10%), and test set (~10%) were first divided based on single-phase images, and multi-temporal images were then assigned to corresponding subsets according to their spatial location. This dataset can support research on levee-like feature extraction, remote sensing change detection, and cross-region model transfer learning.