This publication provides datasets employed in training and testing classification models and a semantic segmentation model designed for quantifying hurricane-induced damage to residential roof subassemblies. These datasets specifically characterize residential properties in Calcasieu Parish, Louisiana. For classification models, images are organized into their ground truth classes and placed in their respective train/test subdirectories. For our segmentation data, each subdirectory contains a collection of images (in .jpeg or .png format), corresponding ground truth masks (.png), and a metadata file (.csv). These resources can be reused for training damage detection models or can be integrated with custom, compatible datasets. Datasets depict damage to residential homes from Hurricane Laura. Raw training image data was sourced from NOAA's National Geodetic Survey's emergency response imagery database while test data was curated from NOAA and the Calcasieu Parish police jury's GIS portal. Ground truth annotation masks were created using the open-source Label Studio software.