Quantifying Uncertainty - talk for the TIES session at JSM 2023 in Toronto.
Advancing environmental statistics through online collaborative groups Chair: Won Chang, University of Cincinnati
Monday, Aug 7: 8:30 AM - 10:20 AM
Session 1831
Topic-Contributed Paper Session
Main Sponsor Section on Statistics and the Environment
Presentations
Uncertainty Quantification for Machine Learning Models in Environmental Systems The increased flexibility of data-driven machine learning models in capturing complex patterns in environmental systems comes at the price of an uncertainty which needs to be quantified. While the topic of uncertainty in machine learning is increasingly acknowledged as necessary to bring more informative predictions, so far only a small number of methods have emerged to address this topic. Our TIES working group composed of statisticians and environmental scientists from around the world investigated this topic, and in this talk we will review the primary outcomes of this exploration. We will present results for traditional time-series modelling techniques as well as for machine learning approaches and show how these frequently deemed "black-box" models have the ability to achieve proper uncertainty quantification with the most up-to-date methodologies.
Wesley Burr, Trent University
Exploration of model agnostic explainability methods for predicting environmental data The International Environmetrics Society (TIES) hosted a working group this past year exploring artificial intelligence in environmental science. This presentation will focus on the group's findings and insights into model agnostic explainable methods on predicting a soil moisture data set from sea surface temperature anomalies. We compare local method feature importance such as Shapley and LIME to global methods such as feature perturbation to explore information that can be gained from these newer methodologies.
Susan Simmons, North Carolina State University
So many methods for multivariate spatial analysis, which should I use? Our TIES working group considers the multivariate spatial outcome problem, specifically the Exact, NNGP and Tensor Product approaches. We explore via a large simulation study which methods should be used based on sample size and the magnitude of correlation between two outcomes. Guidelines for which method is preferable based on predictive performance are given. The methods are illustrated on several datasets with two correlated outcomes.
Edward Boone, Virginia Commonwealth University
Building integrative capacity to supportonline collaboration in diverse multicultural environment The value of applying team science principles in research and community building projects is well recognized by academic institutions, funding agencies and practice communities. We applied the team science principles inseveral NSF-funded projects and seed grants, including the ongoing project related to the development of FamineArchive and establishing data hackathons as a research tool for data scientists and domain experts. We proposedto adapt strategies for building collaborative capacity based on Team Science Toolkit. Using team scienceapproaches for establishing online collaborative groups, we identifi ed several common barriers that reduce long-term motivation and productivity. We also examined the pathways for interdisciplinary knowledge exchange inmultidisciplinary, multicultural, and multigenerational environment. We proposed solutions for tackling these problems by adopting concepts of equity, empowerment, and intellectual humility as guiding principles to acquire,build, and improve integrative capacity for collaboration. We also explored the common features and dynamics ofdata hackathons and identifi ed several pathways to convert them into a powerful research tool.
Elena Naumova, Tufts University
A Bayesian Multisource Fusion Model for Spatiotemporal PM2.5 and NO2 Data in an Urban Setting Following the WHO's new global Air Quality Guidelines, increased pressure is placed on international governments to report adherence to the reduced air pollution concentration thresholds. However, data from ground monitoring stations are sparse, and do not capture the fine-resolution trends.
Here we propose a novel approach to estimate and predict monthly concentrations of PM2.5 and NO2 across London, from 2005 to 2019. This Bayesian assimilation approach calibrates satellite-derived aerosol optical depth and tropospheric column NO2 to ground-monitored concentration data and uses the Pollution Climate Mapping numerical model along with meteorological, population, and atmospheric covariates for prediction across the entire domain.
The model formulation is based on the integrated nested Laplace approximation approach and uses the Stochastic Partial Differential Equation with a non-separable space-time covariance structure and time-varying coefficients. Cross-validation results support the use and predictive power of this model for estimated PM2.5 and NO2 within sparsely-monitored locations at a monthly resolution, compared to competitive models and single source approaches.
Abi Riley, Imperial College London