Identifying Wildland Fire Risks in California Using Machine Learning and Multi-Source Observation Data

Abstract

Climate change has become an urgent issue, posing a severe threat to the ecosystem, public health, and general safety. Wildfires in the state of California are responsible for damages ranging anywhere from a few hundred to hundreds of thousands of acres of burned land. The onset of wildfires near residential areas poses an enormous risk to residents. Better understanding of wildfire patterns and risks are crucial for improved safety precautions and awareness. This study exploits machine learning technology to identify fire risks in California with integrated data from multiple sources, including satellite remote sensing, ground observations, modeling and reanalysis. The Monitoring Trends in Burn Severity (MTBS) Fire Occurrence Dataset, the NASA MODIS Normalized Difference Vegetation Index (NDVI) products, the United States Drought Monitor (USDM) drought index, and temperature and precipitation data from NOAA NCEI, are preprocessed and aggregated to the county level. Then, five variables, including monthly air temperature, monthly air temperature anomaly, drought index, vegetation index, and monthly precipitation, are used to train a neural network model with 4 layers, using a rectified linear unit (ReLU) activation function and a softmax output activation function. Both training and validation demonstrate good accuracy in fire risk assessment, 73.11% in training, and 73.65% in testing. Further analysis and examination show temperature as the most influential component among the variables selected, followed closely by precipitation. This investigation provides potential approaches to improve wildland fire prediction in the state of California.