Estimating Causal Effects of Greenland Blocking on Arctic Sea Ice Melt using Deep Learning Technique

Over the recent decades, Earth scientists have noted a more pronounced shift in climate patterns near the polar regions, specifically the Arctic, in comparison to the rest of the Earth. The increased warming is largely attributed to the diminishing ice cover in the Arctic, which causes solar radiation to be absorbed instead of being reflected by the previously ice-covered areas. The higher ice melt further threatens the wildlife and indigenous communities. To quantify the impact of Arctic warming, also known as Arctic amplification, environmentalists and domain experts rely greatly on dynamic forecasting systems such as coupled Earth System Models. However, attempting to deduce the causal impact of atmospheric processes on the melting of sea ice through dynamical modeling can be computationally expensive. We therefore use a data-driven causal inference method to analyze the causal effects of atmospheric processes on Arctic amplification.

For data-driven Causal Inference (CI) approaches, there are two main categories widely adopted by researchers, namely, Rubin’s potential outcome framework and Pearl’s do-calculus. The first category, potential outcome framework, relies on hypothetical interventions such that it defines the causal effect as the difference between the outcomes that would be observed with and without exposure to the intervention. Whereas the second approach, i.e. do-calculus, identifies causal effects in non-parametric models using conditional probabilities. In either of these approaches, most inference methods are developed for independent and identically distributed (i.i.d) data and rely on linear regression based techniques. Such techniques are unsuitable for time series data and are susceptible to bias due to time-varying confoundedness. Furthermore, the intricate non-linearities inherent in Earth science data make it infeasible to perform causal inference using existing linear regression or marginal structural techniques. To address these intricate challenges, we follow Rubin’s potential outcome framework and introduce TCINet — a novel Time-series Causal Inference Network designed to quantify causal effects within continuous treatment scenarios by leveraging recurrent neural networks with a probabilistic balancing approach. More specifically, we design a deep neural network based potential outcome model by leveraging the long-short-term-memory (LSTM) layers for time-delayed factual and counterfactual predictions with a custom weighted loss. To tackle the confounding bias, we experimented with multiple balancing strategies, namely TCINet with inverse probability weighting (IPTW), TCINet with stabilized weights using the Gaussian Mixture Model (GMMs) and TCINet without any balancing technique. To gain trust in the model performance, we evaluated the model on a synthetic time series dataset comprising a single treatment (cause), time-delayed potential outcome (effect) and two time-varying covariates.

Pursuant to our research on Arctic Amplification, we used TCINet to analyze the causal relation between Greenland blocking and sea ice melt during the summer months of June, July and August (JJA) over the period of 2003-2018. Along with the sea ice observations from NSIDC, the dataset comprised a total of eight atmospheric and oceanic variables from ERA5 reanalysis product including the potential cause, i.e., the Greenland blocking index. We utilized TCINet to predict counterfactual values of sea ice extent over the Barents and Kara Seas after a lead time of eight days by perturbing summer values of GBI to increase by a multiplicative factor of their 40-year trend, i.e., 2 times the 40-year summer trend, 3 times the 40-year trend and so on. Our model predicts that the average daily sea ice extent value in JJA summer months would have decreased by 0.64, 0.65 and 0.69 million square kilometers between 2003 to 2018, given the GBI was increased by 8%, 12% and 16% respectively, i.e. 2, 3 and 4 times the daily trend. Our findings align with the literature on "increasing Greenland blocking index leads to decreasing sea ice extent", further demonstrating our approach’s potential to substantially quantify the impact of key contributors to the melting of Arctic sea ice.

With the aid of TCINet, we highlight the ability of machine learning to make inference using observed data patterns even with limited understanding of the complex underlying processes. We demonstrate how data-driven interventional techniques offer a light-weight and cost effective alternative to traditional dynamical modeling, along with a quantification of associated levels of model uncertainty. Through this work, we wish to pave paths for advancing causal inference in the realm of observational Earth science, and explore the potential of neural networks in inferring climate change in the presence of temporal and spatial confounders.