Fast-track development of an automated solar photovoltaic module detecting framework utilizing open-access multispectral satellite imagery

The southwest coast of Taiwan is an important habitat for migratory waterbirds. However, the use of fossil fuels has led to climate crises. As society's awareness of environmental issues increases, renewable energy including solar power has become an irreplaceable choice. Monitoring the spread of solar photovoltaic (PV) to mitigate its ecological impact has become an important task. Yet, acquiring information related to the geographic information of this renewable energy is often challenging. Remote sensing techniques for identifying the distribution of solar PV modules have been extensively discussed and published in recent years. However, the use of high-resolution imagery, powerful computational hardware, or high-precision but technically challenging frameworks, particularly in terms of computational costs, hinders the rapid updating and dissemination of information on solar PV distribution. The aim of this study is to rapidly develop a PVs detecting framework. To do so, by utilizing the spectral reflectance characteristics of PV modules, along with open-access Sentinel-2 satellite datasets, we selected false color images derived from Bands 8 (842 nm), 11 (1,610 nm), and 12 (2,190 nm). With the foundation of ArcGIS and the U-Net deep learning architecture, we developed a framework with low training costs, simple data processing, and short training time. Through syntheses of false color with different combinations of spectral bands, we achieved a PVs detecting framework with 94.5% accuracy, 0.94 F1 score, and 0.89 Kappa in aquaculture landscape. In our application in the southwest coast region, the overall accuracy reached 96%, with a 0.92 F1 score and 0.9 Kappa performance. This framework will facilitate the rapid monitoring of the spatial and temporal distribution of PV in Taiwan, accelerate social communication, and alleviate ecological impacts.