To achieve carbon neutrality (CN) by 2050, efforts have primarily focused on CN at the level of individual buildings. However, it is essential to consider how CN can be achieved for entire cities, including these buildings. This study aims to estimate the available space for solar panel installation on buildings in urban areas to enhance the utilization of photovoltaic systems. In this study, we propose a semantic segmentation method for estimating building shape (protruding parts and depressed parts relative to the reference plane), which significantly influences the available space for solar panel installation. We conducted semantic segmentation under five different conditions, varying the training and testing datasets to analyze the impact of roads in aerial imagery. Furthermore, to improve the extraction accuracy of protruding parts, this study proposes a weighted overlay method, which overlays five different inference images using various weight combinations. For protruding part extraction using the semantic segmentation method, an F-Score of 30.5% was achieved when roads were fully retained during training and inference, whereas an F-Score of 41.4% was obtained when roads were entirely removed during training and inference. However, the extraction of depressed parts was not successful. Additionally, the weighted overlay method improved the F-Score, achieving a maximum value of 72.5%.