A framework to estimate generating capacities of PV systems using satellite imagery segmentation

The growing interest in global warming has led to various efforts to rely on new renewable energy such as solar power. More accurate energy generation calculation is on-demand with the rise of publicly accessible residential photovoltaic (PV) panels. PV panel area computation is the first step towards the image-based estimation of energy generation from the residential solar arrays. Satellite image segmentation provides a low-cost and simple solution to calculate solar panel area installed on rooftops and the ground. The most relevant approach is to acquire high-quality labels rather than network architectures in many scenarios. This research aims to estimate potential solar energy from satellite images using deep learning segmentation techniques. We compared five architectures such as UNet, UNet++, Feature Pyramid Network (FPN), Pyramid Scene Parsing Network (PSPNet), and DeepLabV3, with different encoders. We converted segmented image pixels into square meters regarding the average slope of panels. We considered the average solar panel efficiency, global tilted irradiation, and the loss coefficient to estimate the annual energy output from the calculated area. UNet with the EfficientNetB7 encoder performed the best with a dice similarity coefficient of 0.9982 for segmentation and a Mean Absolute Percentage Error (MAPE) of 3.5 for energy output estimation.