A large-scale ultra-high-resolution segmentation dataset augmentation framework for photovoltaic panels in photovoltaic power plants based on priori knowledge

Abstract

Most current efforts to improve model accuracy focus primarily on refining the model itself, often overlooking the critical role of dataset quality—particularly in the context of remote sensing big data. Many large-scale extraction studies of photovoltaics (PV) tend to focus on coarse delineation of PV plant boundaries, which limits the potential for more detailed downstream analysis. This paper presents a framework that targets the fine-grained extraction of PV panels within PV power plants, rather than merely capturing the external contours of the plants. By focusing on individual panel-level segmentation, this approach enables more accurate assessments for downstream applications, such as energy yield estimation and spatial optimization. The framework integrates prior knowledge to address challenges posed by land cover, imaging conditions, and background interference. An innovative label correction model reduces pixel-level labeling effort by 75 %, resulting in a more refined dataset. Experimental results show a significant accuracy improvement—from 78 % to 92 %—which is attributed not only to the model refinement but also to the enriched dataset. This dataset augmentation offers substantial advantages for PV mapping, enhancing the precision of energy-related analyses and facilitating more effective solar energy management.