Machine Learning Approach for Enhanced Understanding of California’s Wildfire Smoke

Wildfires are one of the biggest disasters affecting humankind, given the damage to ecosystems, infrastructure, and health, caused by smoke emissions. While the frequency and intensity of wildfires have increased over the past decades due to human activities and climate change, it is still challenging for even sophisticated physical models to accurately project wildfire emissions and predict their transport to downwind areas, because of the complications of physical and chemical processes involved in fire emissions and their interaction with the ambient environment. Moreover, this method is computationally expensive, and numerous complex fire processes parameterized in models are poorly understood, which leads to a reduction in the accuracy of the model outputs. As an alternative method, Machine Learning (ML) algorithms can identify nonlinear relationships between large datasets as numerous input variables and predict future outcomes that might be difficult or impossible to discern using traditional methods.

This study aims to apply ML techniques to explore the data science behind California’s wildfire smoke emissions. We will apply an ensemble of remotely sensed observations and the historical Modern-Era Retrospective Analysis for Research and Applications, the second version (MERRA-2) reanalysis data from 1984 to 2022 to develop a predictive system using various ML algorithms with the goal of finding efficient ways to improve the predictability of smoke emissions. Incorporating fire-related atmospheric and land surface variables, including temperature, wind speed, relative humidity, burnt area, soil moisture, Normalized Difference Vegetation Index, Evaporative Demand Drought Index, evapotranspiration, surface precipitation index, Vapor Pressure Deficit, and the Palmer Drought Severity Index, is essential for employing ML models to estimate wildfire smoke concentration and their transport to receptor areas. The ultimate goal of this study is to apply our ML model results to gain further insight into the complex mechanisms of wildfire smoke emissions and their transport to downwind areas.