The dataset consists of 3617 GeoTIFF images, clipped by a buffer of 2 m around the roof outline with a mask around it, stored in four different folders by roof type: flat, gable, complex and bug. The bugs category includes all images which do not represent buildings, such as construction sites, unclear images, small parts of roofs or simply impossible to recognize with a human eye shapes. The orthophoto used for preparing the dataset is in TIFF format. It was obtained in 2020 through aerial photography with an ultra-wide range digital camera (UltraCam Eаgle Mark 3). The orthophoto has the following characteristics: Height of flight above the terrain: 2850-3200 m; Longitudinal overlap: 60%; Transverse overlap: 30%; Aerial imaged area – 1961 sq. km, of which 1342 sq. km is the territory of the Metropolitan Municipality Sofia. For this project, the study area of district Lozents is 9.2 sq. km; Resolution: 10 cm/pix for the urban area; Bands: RGBA numberer of tiles (georeferenced JPG files): 39. Other applications of the dataset, in addition to rooftop detection and classification, are as follows:: roof recognition model, distinguishing roofs from other urban objects such as streets, trees, cars, etc.; roof segmentation model; recognition and classification of roof elements (chimneys, skylights, dormers, terrace, antennas, solar panels etc.); recognition and classification of roof materials (tiles, metal, asphalt, wood, vinyl, etc.); roof solar potential analysis (the characteristics of the roof regarding the requirements for installation of solar panels).