This proposal develops novel computational and statistical models to address a science question in marine ecology: How will marine organisms adapt and survive under extreme climate stressors, in particular, rising ocean temperatures and their extremes? Addressing this marine ecology question requires prediction of extreme climate temperature variables at scales of one to a few meters; whereas, current global climate models only yield credible insights at 100 kilometers. Key to addressing these science questions is to develop computational models for discovering associations and predictive models from nonlinear and relatively non-stationary systems, where the dependence structures can be complex in space and time. In this project, we propose novel statistical dependence measures that capture nonlinear dependencies and non-stationary properties common in extremes and spatio-temporal applications. In particular, we investigate dependence measures based on copulas that satisfy the equitability property (a new concept in statistics describing measures that are invariant to transformations) and develop computational models that utilize this dependence measure to perform feature selection to identify relevant variables and remove redundant ones on high-dimensional climate and marine ecology data. We then develop novel prediction models, leveraging on advances in sparse models, Bayesian nonparametrics, and knowledge of the physics and science of climate and marine ecology.
All the novel computational methods on feature selection and prediction will enable the discovery of associations and prediction of climate extremes at finer resolutions relevant for marine ecology survivorship prediction. Besides broader impact to society through better marine ecology prediction models, we also provide broader impact to education by leveraging our multi-disciplinary team in offering cross-discipline education and encouraging mentoring of women and minority students into our research program.
