Forecasting next day wildfires through image segmentation using convolutional neural networks
Abstract
In this study, we trained a convolutional neural network based on the UNET architecture to predict the locations of wildfires on the next day through image segmentation. The Residual UNET does this by generating wildfire predictors from the inputs using feature extraction. The open-source dataset we used was curated for predicting next day wildfires and contains remotely sensed images of 12 input features relevant to the task. We train the models using focal Tversky, weighted cross entropy, and dice loss functions and then compared their performance using dice coefficient, recall, and precision. Analyzing the predicted fire masks showed that the Residual UNET models are able to predict the general locations of next day wildfires.
