This project is motivated by a science question in marine ecology: How will marine organisms adapt and survive under extreme climate stressors, specifically, rising ocean temperatures and their extremes? Addressing this challenge requires new developments in computational and data-driven representations and models of marine ecosystems, as well as novel machine learning (ML) methods for translating coarse resolution climate simulations to information relevant at the scale of marine ecology, both of which in turn motivate novel computational and data science methods in spatio-temporal associations and predictive models.

Intellectual Merit: 

We developed: (1) novel methods capable of discovering associations from nonlinear and relatively nonstationary systems; (2) predictive models motivated by methods in ML and computer vision for statistical downscaling; and (3) predictive models for the effect of climate change on marine ecosystems.

We have introduced a new statistical concept of robust-equitability and proposed a practical consistent estimator for a robust-equitable dependence measure. Robust-equitable measures are able to reflect the strength of the deterministic signal in data with continuous independent noise, regardless of the relationship form (linear or nonlinear). This property makes our dependence measure suitable for feature selection and for discovering associations (such as, nonlinear teleconnections in climate). This research led to publications in machine learning, mathematics, and climate informatics.

In general, our approaches will help solve outstanding research challenges in climate and adaptation science. For instance, Global Climate Models (GCMs) currently provide relatively coarse resolution outputs that preclude their application to accurately assess the effects of climate change on finer regional scale events relevant to marine ecosystems. Our team received the Runner-Up Best Paper Award for the Applied Data Science Track at the 2017 ACM Knowledge Discovery and Data mining (KDD 2017: one of the most selective and prestigious among data science and AI peer-reviewed venues) conference for a paper on Deep Learning methods adapted to statistical downscaling. This KDD 2017 paper and a follow-up paper in KDD 2018 on Bayesian Deep Learning, along with an abridged version of the KDD 2017 paper in the International Joint Conference on Artificial Intelligence (IJCAI 2018), led to several high-profile citations and highlights, including in editorial commentaries in the journal Nature as well as a Perspective article also in the journal Nature on combining process models with deep learning in earth systems. Our work on spatial covariance estimation with a rate-optimal double tapering covariance estimator for spatio-temporal data, such as climate data, won the ACML 2017 Best Student Paper Runner-Up Award. Overall, these methods can be leveraged in climate science and network science to advance our ability to both predict and adapt to climate-mediated changes.

We have developed several prediction models for studying the effect of climate change on marine ecosystems resulting in several published works in ecology.  We are using a combination of statistical analyses of climate data and ecological modeling in order to understand the effects of climate change on metacommunities. Our analyses of the CMIP5 climate data has shown that the spatial and temporal autocorrelation of air temperature is expected to increase under climate change (RCP8.5), and that the trend is expected to accelerate after 2040. This increase in the spatial and temporal autocorrelation of temperature is expected to increase extinction risk in ecological systems at both local and regional scales.

In addition to this work, we published several studies highlighting the link between climate change and marine ecology, as well as the implications for spatial management and conservation. Specifically, we (i) identified degree of isolation and ecosystem size as predictors of temporal stability and persistence in multitrophic communities experiencing environmental change, (ii) demonstrated how the relationship between spatial synchrony and temporal stability depends on the degree of environmentally-mediated variation in dispersal, and (iii) showed how the effects of intraspecific diversity are likely to destabilize food webs in a changing world.

Broader Impacts:

This project allowed us to mentor and educate a total of 12 graduate students (four of which are female) from four different disciplines (marine ecology, climate science, machine learning, math). The methods developed in this proposal will most likely generalize to other complex spatiotemporal dynamical systems with suitable modifications. Moreover, we have been active in the community, with leadership roles in the Climate Ready Boston Report and an award-winning work for the Town of Brookline under the aegis of the American Geophysical Union Thriving Earth Exchange. We have also developed innovative study abroad programs and been leading Undergraduate students across disciplines into a Dialogue of Civilization program abroad on Climate Change Science and Policy. Ganguly led the AI section of the Independent Advisory Committee for Applied Climate Assessment (IAC). This report, once published, was highlighted in the national and international mainstream media. 

Last Modified: 11/27/2020
Modified by: Jennifer G Dy