Hurricanes can cause massive power outages and pose significant disruptions to society. Accurately monitoring hurricane power outages will improve predictive models and guide disaster emergency management. However, many challenges exist in obtaining high-quality data on hurricane power outages. We systematically evaluated machine learning (ML) approaches to reconstruct historical hurricane power outages based on high-resolution (1 km) satellite night light observations from the Defense Meteorological Satellite Program (DMSP) and other ancillary information. We found that the two-step hybrid model significantly improved model prediction performance by capturing a substantial portion of the uncertainty in the zero-inflated data. In general, the classification and regression tree-based machine learning models (XGBoost and random forest) demonstrated better performance than the logistic and CNN models in both binary classification and regression models. For example, the xgb+xgb model has 14% less RMSE than the log+cnn model, and the R-squared value is 25 times larger. The Interpretable ML (SHAP value) identified geographic locations, population, and stable and hurricane night light values as important variables in the XGBoost power outage model. These variables also exhibit meaningful physical relationships with power outages. Our study lays the groundwork for monitoring power outages caused by natural disasters using satellite data and machine learning (ML) approaches. Future work should aim to improve the accuracy of power outage estimations and incorporate more hurricanes from the recently available Visible Infrared Imaging Radiometer Suite (VIIRS) Day/Night Band (DNB) data.
