Deep learning for short-term prediction of wildfire using geostationary satellite observations

Abstract

The aim of this thesis is to utilize the Geostationary Operational Environmental Satellite (GOES) data for predictions regarding the intensity and potential path of wildfires. Using GOES to identify wildfires and extracting data from those events to help train a deep learning model. Three fires were selected for training the deep learning model: the Sequoia, Calwood, and Maui fires. The GOES data of the fires was obtained from band 7 which operates in the Shortwave Window or 3.9μm wavelength, band 7 is able to capture hotspots which is beneficial for wildfire prediction. The radiance data from band 7 is pulled from an Amazon Web Service (AWS) and becomes part of a dataset of 2513 samples. The data is then stacked to form a time series of approximately two hours and converted into a compressed h5 file. The pipeline distributes the dataset by taking in twenty five minutes of input data and feeding four different models to predict seventy five minutes, one hundred minutes, and one hundred and twenty five minutes of data. The data is then fed into a deep learning model utilizing a model known as Self Attention Gated Recurrent Unit (SaGRU). The SaGRU is tested four times, once for predicting seventy five minutes, once for predicting one hundred minutes, and twice for one hundred and twenty five minutes. The models were then compared against each other regarding Mean Squared Error (MSE) and Mean Absolute Error (MAE) along with the Normalized Mean Squared Error (NME) and the Normalized Mean Absolute Error (NMAE). Each metric was taken along multiple thresholds comparing the performance when hotspots are present and when hotspots are absent. The resultant showed that regardless of the sequence length, there was minimal negative impact on early predictions, but as the predicted sequence increased significant loss could be seen on the later predicted frames.