[A-CG46] Biogeochemical Cycles in Land Ecosystem

[ACG46-P11] Construction of a dataset for creating a photovoltaic panel distribution map in Japan

*Haru Ueda¹, Wei Yang^1,2 (1.Graduated school of science and engineering, Chiba University, 2.Center for Environmental Remote Sensing, Chiba University)

Keywords:Satellite Remote Sensing, Deep Learning, Photovoltaic Panels

Photovoltaic (PV) power generation is a key renewable energy source; however, concerns about the environmental impact of PV installations have led Japan’s Ministry of the Environment to establish guidelines for their deployment. Therefore, it is crucial to understand the spatial and temporal distribution of PV panels across Japan. While PV panels can be effectively detected using high-resolution satellite or aerial imagery and deep learning algorithms (e.g., convolutional neural networks, CNNs), supervised deep learning models require extensive data annotation, making large-scale mapping costly and time-consuming. Consequently, this study aims to propose a semi-automated dataset construction approach using the latest Segment Anything Model (SAM) to reduce costs and accelerate model development. After identifying the locations of PV panels, SAM can automatically generate polygons for them. As a result, annotation time for 2,500 points was reduced from 200 hours per person to just 12.5 hours per person. This method significantly reduces annotation time, enabling the creation of larger datasets and more comprehensive spatial analyses within limited resources. As a benchmark, we constructed a training dataset covering a 40 km² area in the Inbanuma watershed, Chiba Prefecture, and trained a deep learning model to map the PV panels based on satellite images. The trained model achieved accuracy comparable to or higher than that of existing studies. This demonstrates that the proposed method enables the construction of larger datasets without compromising accuracy.