Extracting 10-m Resolution Photovoltaic Landcover Using A Slightly Modified U-Net and Sentinel-2 Images

Centralized photovoltaic (PV) power plants, as the main form of solar energy utilization, has been world-widely constructing in recent two decades, which makes photovoltaic landcover (PVL) becoming an important landcover type. Accurately mapping PVL provides basis for both the estimation of grid-connected PV power capacity and the optimization of spatial layouts of PV power plants. However, mapping PVL at large-scale is still facing challenges due to the complexity of landcover with PV panels, and the differences of PV panel sizes and arrangement patterns. This study aims to apply a slightly modified U-Net model for precisely extracting PVL from 10-m resolution Sentine1-2 (S2) images. Firstly, combining visual interpretation and historical data of PV power plants, we selected 16 sample sites and 4 test sites for study, the sample sites are scattered in China, while the test sites are typical of three centralized PVL types. Secondly, the S2 images in 2022 in these 20 sites were labeled and cropped into chips with 64 * 64 pixels. Thirdly, the modified model was trained and validated using the sample dataset with four schemas of channel combinations. Finally, the trained model was adopted for mapping PVL in the 4 test sites. The validating results show that the slightly modified U-Net achieves excellent performance with the mean intersection over union (MIoU) of the four schemas all above 97%. Visually interpreting the testing results also demonstrates that the modified model competes to mapping PVL at large-scale.