Hurricane Damage Prediction on Satellite Imagery based on Neural Networks

Abstract

Accurate categorizations of Hurricane damage on specific locations could significantly facilitate rescuing teams and rescuing resources to be deployed to where they are needed the most. In addition, it could aid analysts to predict further predict the potential damages incoming Hurricanes could bring to various locations based on previous categorizations of satellite images about damaged terrains and buildings. This study possesses a Neural-network-based prediction model, in which CNN and FFNN of various parameters and structures are performed to make predictions regarding if a certain location is damaged due to a Hurricane from images captured by Satellites. Based on various aspects of the overall performance of models, including accuracy, AUC score, Loss curve, confusion matrix, F1 score, model fitting time and training time, the best model regarding this task is AlexNet with an accuracy of 96.77% and F1 score of 0.9816 despite its slightly longer training time of 63s per epoch. The results of fellow neural network models suggest that neural network models are capable of handling images categorization and prediction problems regarding satellite images of Hurricane, thus helping optimize resources expenditure and improve efficiency for further related analyzes.