Predicting Wildfire using Live Fuel Moisture Content with Machine Learning

Forest fires are responsible for significant losses to societies, ecosystems, and economies worldwide. To keep a check on these losses and to avoid wildfires, predicting the occurrence of these forest fires are important because these can help in wildfire management and prevention. In recent times, with the availability of real time high resolution remote sensing satellite data over larger geographies, it is becoming easier to study geologic features of our planet. One such example is the live fuel moisture content (LFMC), which is defined as the "mass of vegetation water per unit dry biomass", a crucial factor determinant of wildfire risk. Using the Live Fuel Moisture Content in combination with ground truth wildfire data from the United States Geological Survey, I predict whether there will be a fire in the near future within a 8 km grid over the area of interest using supervised machine learning models such as KNN, SVM, Random Forest, and Neural Networks. We conducted two experiments, using a single image within 15 days of a fire, and a combination of 3 images within 45 days of a fire. Our results show that the random forest performs best with an accuracy of 76.06% for a single image, and 75.07% for multiple images.