Statistics Seminar

Abstract: In this talk, we will present two papers that incorporate and revisit ideas proposed in the Bayesian nonparametrics literature in a spatial context to address problems in Earth Sciences. 

Namely, in the first work we will consider soil moisture, that is, the amount of water stored in soil pores, a factor of critical importance in terrestrial hydrology, influencing land processes such as crop health, wildfire dynamics and hydrologic extremes. Despite its undeniable relevance, soil moisture data is not widely available, with few in-situ probes providing localized data and more extensive spatial information provided by satellites at a lower temporal frequency. To encourage the collection of new in-situ data and provide guidance for future sampling campaigns, we propose a statistical model that aims to learn about the scale of dependence and variability in the spatial process from observed data. The model combines the Multi-Resolution Approximation idea of Katzfuss (2017) with the CUSP prior of Legramanti et al. (2020) to identify regions in a spatial domain where a process displays similar levels of spatial variation. 

In the second project, the focus is instead on gaining useful information that can be leveraged to design air pollution mitigation strategies. Specifically, looking at observed concentrations for 6 of the major components of PM2.5 over space and time, we develop a modeling framework that aims to identify the major sources contributing to the pollutants concentration. Our source apportionment model embeds the multiplicative gamma process shrinkage prior of Bhattacharya and Dunson (2011) with a spatial-functional modeling framework thus allowing to infer upon both he number of sources and their local contribution to the observed pollution levels at a given site.