Abstract:In recent years, the country has vigorously developed distributed photovoltaic new energy. The power grid needs advanced planning and reasonable layout in the integration of new energy, therefore, it is necessary to accurately evaluate and analyze the potential of regional rooftop photovoltaics. This article proposes a semantic segmentation based method for evaluating the photovoltaic potential of building roofs. The method uses satellite images from an open platform as the data source, integrates remote sensing technology and artificial intelligence, and identifies and calculates the roof information of buildings in the research area. By establishing a building satellite image dataset, a network architecture framework based on Deeplab v3+was designed, and a CBAM attention mechanism module was added to the recognition framework to focus on and enhance different types of roof features, improve model accuracy and prediction precision, and achieve recognition and photovoltaic resource estimation of different types of roofs such as rural residential buildings, industrial parks, urban high-rise buildings, and installed photovoltaic roofs. Match the identified connected domains with the registered entities on satellite maps, obtain relevant geographic location interest point information of the connected domains, and combine solar module attributes, building attributes, meteorological environment attributes, etc. to classify and evaluate the energy-saving and carbon reduction potential of roofs of buildings that have installed, are suitable for installation, and are not suitable for installation of solar photovoltaics.