• building detection
• computer vision
• deep learning
• satellite imagery
• roof segmentation

Traditional means of obtaining geo-localized information on building equipment (e.g., solar panels and other installations) are limited in availability and geo-spatial resolution and basically rely on intensive human on-site analysis. In this work, a different approach, based on computer vision algorithms and high-resolution color satellite imagery, is investigated for automating this information acquisition process. The main objectives are therefore the creation of a model that allows the automatic recognition of roofs of buildings and the creation of a model that allows the automatic recognition of equipment on roofs in satellite images. The problems that will be analyzed fall in the field of image segmentation, one of the fundamental problems in computer vision; thus some of the techniques proposed in the literature in recent years will be initially studied, in particular the U-Net and Pyramid Attention Network architectures. These architectures will then be applied and their effectiveness will be studied, comparing them with the current state of the art.