Spatiotemporal Forecasting and Causality Methods for the Arctic Amplification

Abstract

The Arctic is a region with unique climate features where warming has been almost three times as fast as the rest of the world. The warming of the Arctic, also known as Arctic amplification, is led by several atmospheric and oceanic drivers. One of the devastating implications of Arctic amplification is the unprecedented sea ice loss as demonstrated by satellite observations. However, understanding the causes of sea-ice variations and its feedback on the atmospheric processes is a complex task. This brings us to three key research areas studied in this dissertation: (i) accurate forecasting of climate data in the Arctic on sub-seasonal to seasonal scales, (ii) estimating the influence of atmospheric processes and their time-varying effects on the sea-ice and ice-sheet variations. (iii) estimating the spatial interference of atmospheric processes and their time-varying effects on the sub-regional sea-ice variations. For the first topic, this research explores the potential of data-driven approaches to study sea ice variations by proposing custom deep learning modeling techniques to learn spatiotemporal variations in the sea ice. The proposed models reduce the prediction error by 60% as compared to the state-of-the-art approaches, while the work further contributes to accurate long-term forecasting beyond the seasonal barrier. For the second topic, the research utilizes potential outcome framework to infer causation in climate data on daily and monthly temporal scales using custom models based on recurrent neural networks to infer causation under continuous treatment, and a novel probabilistic balancing technique to reduce the confounding bias. For the third topic, this work formalizes the concept of spatial interference in case of time-varying treatment assignments. By extending the potential outcome framework and utilizing latent factor modeling to reduce the bias due to time-varying confounding, the framework leverages the power of U-Net architecture to capture global and local spatial interference in data over time. Being the first of its kind deep learning based spatiotemporal causal inference technique, the proposed approach shows advantages over several baseline methods based on the experiment results on two synthetic datasets, with and without spatial interference. The empirical results on Arctic dataset not only align with domain knowledge, but also pave paths in quantifying the impact of causal drivers of climate change in the Arctic. Overall, these predictive and inferential models have the potential to generalize for multiple downstream tasks and can be extended to other domains beyond Earth Science.