Classification of post-wildfire aerial imagery using convolutional neural networks : A study of machine learning and resampling techniques to assist post-wildfire efforts

Sammanfattning: Assessment of post-wildfire damages to human structures is a manual task which currently uses ground-level observations of the structures by a human inspector to classify the burn severity of the affected structure. This study investigated the potential of using machine learning and specifically computer vision techniques in order to produce a classification of burn severity using post-wildfire aerial imagery. Specifically, a convolutional neural network model was trained on post-wildfire aerial imagery of affected human structures, and learned to classify their burn severity. The dataset used also suffered from a class-imbalance problem, meaning that the ratio of the different burn severity classes was skewed. Resampling techniques were studied as a method of improving performance given the class imbalance problem. The study showed that convolutional neural networks were able to provide valuable classifications of the affected structures given post-wildfire aerial imagery. However, the results of the study showed that resampling using random oversampling did not provide an increase in model performance, and in fact lead to a worse performance when compared to a model trained on the same dataset without resampling.