Statistical methods for the interpretation, prediction, and localization of remotely sensed atmospheric pollutants

Abstract

We present two statistical modeling efforts that seek to address pressing environmental issues through the use of remotely sensed data. The first study is motivated by the extreme fire seasons now commonly experienced across the globe (e.g. the 2015 Indonesian forest fires and the 2019/2020 Australian bush fires). We develop interpretable models for remotely sensed carbon monoxide, a proxy for fire intensity in the Southern Hemisphere, fit using a flexible regularization framework. These models are parsimonious by design, allowing for scientific insight into the primary climate drivers of fire season intensity in different regions. The models have good predictive skill at considerable lead times, making them a useful tool for predicting upcoming fire season intensity. The second study is motivated by a growing dependence on natural gas for energy in the United States. Methane (the primary component of natural gas) burns cleaner than coal and oil but is a potent greenhouse gas. Therefore, limiting emissions during natural gas production is essential if it is to be considered a cleaner alternative to other fossil fuels. With the goal of localizing small-scale emissions, we develop a hierarchical spatial model for estimating methane concentrations on a fine grid given coarsely pixelated satellite observations. We apply our model to a satellite overpass of the Denver-Julesburg (DJ) Basin (located in northeast Colorado) to demonstrate its effectiveness. We use conditional simulation for uncertainty quantification and inferences related to emissions monitoring.