PromptNet: Prompt Learning for Roof Photovoltaic Potential Assessment

Abstract

An increasing number of works have been proposed to use remote sensing images to assess the potential for rooftop Photovoltaic (PV) energy development in buildings. However, most methods focus mainly on the remote sensing images themselves, ignoring the key prior information of building type. Thus most works with Deeplabv3+ as backbone present suboptimal performance. To overcome this challenge, we propose a novel approach PromptNet that embeds the building types as prior knowledge and feed it into prompt learning for predict roof PV energy Potential. Specifically, a pre-trained semantic segmentation network, Deeplabv3+, is first constructed to detect potential building rooftops from remote sensing images. Then, the buildings are categorized into five types based on their functions, including government buildings, public buildings, industrial and commercial factories, agricultural housing, and other building types. Finally, by using prompt learning, the prior knowledge of buildings is established and associated with the rooftops that are suitable for PV energy development. This is embedded into a deep learning network, filtering out unsuitable rooftops, and significantly improving the accuracy of rooftop PV energy development. Comprehensive experiments show that the proposed method achieves 81.18% accuracy and 76.90% IOU in predicting the potential for rooftop PV energy, a 10.97% improvement in IoU compared to the backbone without prior knowledge.