Dataset for paper \A Road Connectivity Enhancement Framework Based on Improved PIDNet and Post-Processing Steps for UAV Remote Sensing Imagery\

The original UAV remote sensing dataset was provided by the Chengdu Land Reclamation and Ecological Restoration Center. It consists of true-color RGB images acquired in May 2023 over Caihua Village, Chongzhou City, Sichuan Province, with a spatial resolution of 0.1 m. The images, covering approximately 2 km\u00b2 at a size of 12,288\u00d78,704 pixels, have undergone geometric correction and denoising.The dataset includes seven land cover classes: Farmland, Woodland, Bare soil, Roads, Building, Water, Other, and Background. To facilitate semantic segmentation, the large images were partitioned into 512\u00d7512 patches using a sliding window approach. The dataset was then randomly split into training, validation, and test sets at a 4:1:1 ratio. Five-fold cross-validation was applied during training on the combined training and validation subsets. To improve model generalization and data diversity, four augmentation techniques\u2014vertical flip, horizontal flip, random rotation, and cutout\u2014were employed, expanding the training samples. Each fold thus contains 2,191 training, 114 validation, and 113 test samples.