Solaris: Deep Learning Based Model for prediction of solar power yield on rooftop using aerial imagery

This paper describes the implementation of Deep learning and computer vision, specifically U-Net Model for estimating the rooftop-areas where solar panels can be installed and calculate the amount of solar power that can be generated. U-Net is a type of neural network which is used for image segmentation. Aerial imagery is used to obtain rooftop surfaces which are available for solar panel installation. It is observed that not all the surfaces are suitable for installing solar panels, thus such rooftops can be discarded. Further, the rooftops which fit are input to the U-Net model for estimating the area where solar panels can be installed. The solar radiation data is another input given to the system. Using these inputs, we can calculate the amount of solar yield from the rooftop surface area after installing the solar panels. The results demonstrate that the model achieves high accuracy in segmenting rooftop areas, with a mean Intersection over Union (IoU) of 0.7627 and a mean Dice Score of 0.8534 in the testing phase.