Wildfire Detection and Perimeter Mapping using Satellite Imagery and Machine Learning with Hyperopt Tuning

In part due to climate change, the last few years have been some of the warmest on record and characterized by hot and dry weather. This led to a frequent outbreak of wildfires, especially in already dry areas such as California. Experiments were conducted to evaluate the ability of machine learning models to detect wildfires and map the areas burnt using satellite images. For the detection of wildfires, machine learning models of different complexities are trained to distinguish between images containing wildfires and images containing no wildfires. The tested models achieved consistently training accuracies above 90% and testing accuracies above 70%. HyperOpt was then used to fine tune the models’ hyperparameters to improve their accuracy. For mapping the areas burnt by wildfires referred to for the rest of the paper as wildfire perimeter mapping, a preliminary burn map is produced mathematically from each image. The preliminary map is then refined using an object detection model. The refined burn maps achieved an average accuracy of around 10%. However, in a few cases where the original satellite images have high image quality, the refined burn map that was produced reflected the recorded burn area with above 90% accuracy. In conclusion, machine learning models alongside satellite images have the potential to be used for quick and efficient detection of a wildfire outbreak. With some improvements to the current process, machine learning also has the potential to accurately determine the area burnt by the wildfire at any given time simply using a satellite image - a significant improvement over traditional methods such as hand sketching or GPS walk.