Polygon-based mapping of photovoltaic systems and estimation of energy generation potential

Renewable energy is a key component for the goal of carbon neutrality. A significant contribution is made by photovoltaic systems. This work combines machine learning and geographical processing to identify photovoltaic systems in aerial images and generate the corresponding vector data with geographic alignment information. For this, a combination of Inception V3 model for classification and a DeepLab V3+ model for segmentation is used to extract photovoltaic systems. In addition, the power generation potential of photovoltaic systems are predicted by a linear regression model. The used model is applied to two datasets from different data sources for the same study area. The segmentation maps archives in both cases a Intersection over Union for the photovoltaic class of at least 75 %. Based on the identified photovoltaic systems the distribution shows a high number in residential areas whereas the highest energy generation potential is in industrial and commercial areas.