Deep Learning–based LOD-2 Building Modeling and Its Application to Solar Energy Estimation Using Unmanned Aerial Vehicle Imagery

In this paper, the solar energy potential of multiple roof planes is presented. The main objectives of this research are to estimate the annual solar energy potential and the energy generated on individual roof planes using unmanned aerial vehicle imagery. The methodology is based on deep learning for building footprint extraction and roof-plane segmentation using RANdom Sample Consensus (RANSAC). A multi-criteria selection process was used to determine the most suitable roof planes based on several parameters, including surface orientation, slope, solar insolation, and roof size. We implemented our algorithm to determine the optimal placement of photovoltaic panels and to calculate the resulting energy generation. Based on the results obtained, the proposed method is effective for estimating potential solar energy generated on multiple roof planes. Our results indicate that the roof planes produce an annual power generation of 553 219 kWh. These findings underscore the significant potential for reducing carbon emissions through the implementation and expansion of rooftop photovoltaic systems.