Wildfire detection through deep learning based on Himawari-8 satellites platform

The accurate identification of the location, intensity, and spread of wildfires is an essential early-stage precaution for reducing wildfire damage. Satellite imaging platforms, particularly those with high revisiting frequencies and fine spatial resolutions, represent the most efficient possible means of monitoring wildfires dynamically. However, the extraction of accurate fire-related information from satellite images remains challenging, and few studies have investigated the use of remote sensing data from satellites with geostationary orbits. The present work addresses these issues by applying over 5,000 images obtained from the geostationary Himawari-8 satellite of a severe Australian wildfire occurring from November 2019 to February 2020 to train and test a fully connected convolutional neural network (CNN) for identifying the location and intensity of wildfires. The proposed CNN model obtains a detection accuracy greater than 80%, which greatly exceeds that of other machine learning algorithms, such as support vector machine and k-means clustering. Moreover, the CNN model can be trained in a relatively short period, even when employing large training datasets, and predictions can be made in just one or two minutes. The proposed model provides insight into the application of deep learning methodologies for wildfire monitoring based on the imagery provided by geostationary satellites, and support for developing similar satellite missions.