UAS imagery of modified-bitumen flat-roof defects: tiled binary dataset

Tiled and augmented unmanned-aerial-system (UAS) imagery of modified-bitumen flat roofs, labelled for binary defect classification (Defect / No-Defect), supporting the paper "A Tile-Based CNN-SVM Architecture for Autonomous Unmanned Aerial Systems Inspection of Flat Roofs." Full-resolution aerial photographs were divided into 640×360-pixel tiles and labelled by an expert annotator. The data are organised into a photo-level train/validation/test split (75:15:10): every tile and augmentation derived from a given source photograph remains within a single split, and no photograph appears in more than one split. Augmentations are in the training split only; validation and test use original tiles. manifest.csv is the reproducible split index (split, label, photo_id, filename, is_aug). Splitting at the photo level removes tile- and augmentation-level leakage. It does not eliminate roof- or flight-level leakage: consecutive photographs of the same roof may overlap spatially and can fall in different splits, so reported accuracies should be read with that residual effect in mind. Label 0 = Defect (positive class). Contents: 43,383 training, 3,869 validation, 2,540 test images across 5,240 source photographs. License: CC-BY-4.0.