Mapping Meteorological and Topographic Controls on Camp Fire PM2.5 Using Geographically and Temporally Weighted Regression

Wildfires are increasing in frequency, intensity, and duration, and so are their negative consequences. Across the western United States, wildfire smoke has become the leading driver of rising PM_2.5 levels, and large populations are being exposed to this during increasingly frequent smoke events. However, the forces that control the pollution extent and transportation remain partially understood. It is understood that pollution is controlled by the interaction of meteorological, topographic, and boundary-layer conditions, yet this interaction is not fully captured by most existing techniques, particularly conventional regression methods. In regions of complex terrain and meteorology with recurring wildfires, such as California, this becomes a real issue.

This study applied Geographically and Temporally Weighted Regression (GTWR) to examine observed PM_2.5 concentrations during California’s 2018 Camp Fire. GTWR allows regression coefficients to vary across both space and time, making it possible to identify where and when different controls were most strongly associated with PM_2.5. GTWR explained 85.4% of the variance in observed PM_2.5 concentrations, with an RMSE of 15.02 µg m⁻³, and reduced RMSE by 43% relative to multiple linear regression. The local coefficient maps showed that PM_2.5 controls varied strongly across California during the event. Under the Diablo wind regime, the Central Valley appeared to function both as a basin where smoke accumulated under shallow mixing and stagnant conditions, and as a corridor that transported smoke southward. Active-fire-count coefficients were positive near the Camp Fire but negative across much of central and southern California, indicating that PM_2.5 in those regions was driven more by transported smoke from the Camp Fire than by nearby fire activity. A kriging-based spatial consistency check showed strong agreement between the GTWR-estimated PM_2.5 surface and an observation-interpolated PM_2.5 surface.