Spatiotemporal Interpolation of Wildfire Smoke-Derived PM2.5 Using Deep Neural Networks

As the planet warms, wildfires are expected to become a major contribution to air pollution levels, specifically across the western United States (Calvin et al. 2023; Abatzoglou and Williams 2016). The burning of vegetation releases particulate matter–a mixture of solid particles and liquid droplets that hang in the air and can be inhaled (US EPA 2016). Around 90% of this particulate matter is less than 2.5 microns in diameter, referred to as PM2.5 (Vicente et al. 2013). Unfortunately, studies attempting to investigate associations between such health outcomes and wildfiresmoke-derived PM2.5 are plagued by low statistical power and data with poor spatial resolution (Reid et al. 2016). This thesis explores a recently proposed, Deep Learning-enabled interpolation method (Amato et al. 2020) to predict ground-level smoke-derived PM2.5 (SPM2.5) concentrations. I use a AutoEncoder to allow the inclusion of important meteorological covariates, expanding on the previously proposed methodology. The AutoEncoder shows a promising ability to reconstruct SPM2.5, with a post-reconstruction unnormalized MSE of 0.295. However, the interpolation proposed by (Amato et al. 2020) does not handle discontinuous data well, suggesting that further study of machine learning on discontinuous and non-differentiable data is warranted.