Deep Learning for Detecting Tilt Angle and Orientation of Photovoltaic Panels on Satellite Imagery

The goal of this research is to accomplish two tasks that increase the accuracy of the process of estimating solar power generation in real time for different regions around the world. Specifically, we explain a method for detecting the tilt angle and installation orientation of photovoltaic panels on rooftops using satellite imagery only. The method for detecting tilt angles is based on their dependence on the roof shapes. As for the architectures used in this research, we chose MobileNetV2 and Yolov4 since both require only medium hardware resources, without the need for graphics processing units (GPUs). Since it was difficult to find a suitable data set, we had to create our own, which, although not large, was proven to be sufficient to confirm the capabilities of our method. As for the final results, our approach provides good predictions for the tilt angle and the orientation of photovoltaic panels based on a data set of images from six different locations in Europe collected via Google Maps.