Automatically Calculating Hail and Hurricane Damage to Photovoltaic Systems in Satellite Imagery using Deep Learning Techniques

We present a novel methodology for automatically identifying and calculating visible photovoltaic module damage in satellite imagery, following extreme weather events such as hailstorms and hurricanes. To do this, we train a series of deep learning models using imagery data sets taken after a major hailstorm and Hurricane Maria to identify PV modules and associated damages. Our approach for automatically identifying and computing hail-related PV damage in satellite imagery is very robust, achieving an F1-score of 0.97 on test data when identifying imagery containing hail-damaged PV installations. Furthermore, this algorithm auto-computes the percentage of PV damage with a median average error of 0.23%. Our automated methodology for calculating hurricane damage is less robust, achieving an F1-score of 0.72 when identifying hurricane-damaged PV installations. This methodology, which compares solar module masks in pre- and post-hurricane imagery to see how many modules have blown off, is sensitive to error in deep learning mask results, which are fed into an image registration process so pre- and post-storm masks can be directly compared. When ground truth mask inputs are fed into the image registration methodology, damage detection improves significantly, achieving an F1-score of 0.83, and a median average error of 8.5% when estimating percentage PV damage.