Deep Learning Ensemble and Multi-Criteria GIS for High-Fidelity Rooftop Solar Potential Mapping

Abstract Accurately mapping urban rooftop solar potential is essential for cities like Amsterdam that are pursuing net-zero emissions. This study presents an innovative framework that combines high-resolution geospatial data with an advanced deep learning ensemble to identify existing solar panels and untapped suitable rooftop areas. The predictions from a meticulously trained ensemble of deep learning models were integrated using both simple and performance-weighted majority voting. The weighted ensemble achieved an accuracy of 0.95, an F1 score of 0.91, and a Matthews correlation coefficient of 0.88, outperforming individual models. Rooftop suitability was assessed using a multi-criteria approach, which incorporated a high-resolution digital surface model (DSM) to derive slope, aspect, and solar irradiation. A novel solar irradiation model was developed that enhanced the precision of yield estimates by adjusting atmospheric transmissivity and diffuse fraction based on monthly cloud cover data from Amsterdam. This framework provides district-wise spatiotemporal solar irradiation and photovoltaic yield estimates. Based on our model estimates, current installations have a potential of 140 GWh annually, while there is a significant untapped potential of 1276 GWh on suitable rooftops. These detailed insights can help urban planners optimize solar energy deployment and support the city’s carbon neutrality goal by 2050. Graphical Abstract