S3Former: A Deep Learning Approach to High Resolution Solar PV Profiling

Abstract

As the negative impact of climate change escalates, the global necessity to transition to sustainable energy sources becomes increasingly evident. Renewable energies have emerged as a viable solution for users, with Photovoltaic (PV) technology being a favored choice for small installations due to its high reliability, competitive market and increasing efficiency. Accurate mapping of PV installations is crucial in improving grid management, facilitating the integration of renewable energy, encouraging active participation from prosumers, and optimizing the economic performance of decentralized energy markets. To meet this need, S3Former is introduced, which is designed to segment solar panels from aerial imagery and provide size and location information critical for analyzing the impact of such installations on the grid. Although computer vision has become a preferred choice for such implementations, solar panel identification is challenging due to factors such as time-varying weather conditions, different roof characteristics, Ground Sampling Distance (GSD) variations and lack of appropriate initialization weights for optimized training. To tackle these complexities, S3Former features a Masked Attention Mask Transformer incorporating a self-supervised learning pretrained backbone. Specifically, the model leverages low-level and high-level features extracted from the backbone and incorporates an instance query mechanism incorporated on the Transformer architecture to enhance the localization of solar PV installations. Moreover, a self-supervised learning (SSL) phase (pretext task) is introduced to fine-tune the initialization weights on the backbone of S3Former, leading to a noticeable improvement on the results. To rigorously evaluate the performance of S3Former, diverse datasets are utilized, including GGE (France), IGN (France), and USGS (California, USA), across different GSDs. Our extensive experiments consistently demonstrate that the proposed model either matches or surpasses state-of-the-art models (SOTA) and validate the benefit of using the SSL method to improve the segmentation architecture. Source code is available upon acceptance.