A novel weakly-supervised method based on the segment anything model for seamless transition from classification to segmentation: A case study in segmenting latent photovoltaic locations

Abstract

In the quest for large-scale photovoltaic (PV) panel extraction, substantial data volumes are essential, given the demand for sub-meter rooftop PV resolution. This requires the concept of Latent Photovoltaic Locations (LPL) to reduce the scope of the amount of subsequent processing. In order to minimize manual annotation, a pioneering weakly-supervised framework is proposed, which is capable of generating pixel-level annotations for segmentation based on image-level annotations and provides the two datasets required for the classification-then-segmentation strategy without more annotations. The strong noise-resistance of the Segment Anything Model (SAM) is discovered in the extremely difficult rough coarse pseudo-label refinement, which, after integrating a probability updating mechanism, achieves a seamless transition from scene classification to semantic segmentation. The resulting national LPL distribution map, rendered at a 2 m resolution, showcases a commendable 92 % accuracy and a F1-score of 91 %, and the advantages of the framework in terms of efficiency and accuracy have been verified through a large number of experiments. This process explores how to use fundamental large models to accelerate the remote sensing information extraction process, which is crucial in the current trajectory of deep learning in remote sensing. The relevant code is available at https://github.com/Github-YRQ/LPL.

Keywords

• High-resolution remote sensing
• Image-level label
• Large-scale study
• Noisy learning
• Pseudo-label refinement