Hindcasting Fine Particulate Matter in Alaska During Wildfire Seasons

Background<b> Alaska’s frequent wildfire activity affects air quality across large regions, highlighting the need for reliable surface-level fine particulate matter (PM<sub>2.5</sub>) concentrations estimation methods in areas with limited monitoring coverage. </b>Aims<b> We aim to develop a computationally efficient machine learning approach for estimating PM<sub>2.5</sub> concentrations, ensuring model accuracy remains above an R<sup>2</sup> threshold of 0.80. </b>Methods<b> We constructed random forest models using data from May through September 2019–2022. Inputs included land cover and land use, meteorological variables, fire activity metrics, aerosol optical depth (AOD) observations, and measurements from both low-cost and regulatory-grade monitors. </b>Key results<b> All models demonstrated strong explanatory power. The top-performing model, which incorporated AOD, achieved a five-fold cross-validated R<sup>2</sup> of 0.865. Models relying solely on active fire data (excluding AOD) also performed well, with an R2 of 0.833. </b>Conclusions<b> Models based on fire activity metrics, such as total fire count, can reliably estimate regional PM<sub>2.5</sub> levels. Incorporating additional variables like wind patterns yielded only marginal improvements, underscoring the value of streamlined modeling approaches. </b>Implications<b> The approaches presented here can enhance air quality monitoring in rural and fire-prone regions by generating consistent PM<sub>2.5</sub> estimates. With expanded observations and strategic deployment of low-cost sensors, these models could support near-real-time monitoring and long-term wildfire smoke exposure research. </b>