Over the last three decades, the warming of the Arctic sea ice has been almost twice as fast as the rest of the world. This phenomenon, known as Arctic amplification, is caused by several atmospheric and thermodynamic processes. If the amplification continues at the same rate, it is feared that this amplification will further alter the climate patterns across the globe and lead to extreme weather events. Understanding the causes of sea-ice variations and its feedback on the atmospheric processes is a complex task. The first step in studying such a problem is the development of holistic causal graphs that can represent the underlying causal mechanisms of Arctic’s thermodynamic system. Once this graph is discovered, the next step is to estimate the influence of causal effect on a spatiotemporal scale; however, this cannot be accomplished in the presence of confounders varying both, spatially and temporally. This brings us to three key research areas: (i) accurate forecasting of sea ice on sub-seasonal to seasonal scales, (ii) combining results from multiple causal discovery methods to generate holistic aggregated causal graphs (iii) estimating the influence of atmospheric processes and their space-time-varying effects on Arctic sea-ice variations.
 
To cater to these research directions, this work proposes techniques for developing custom deep learning models to learn spatiotemporal variations in sea ice, followed by the development of a potential outcome framework to infer causation in climate data on both spatial and temporal scales using our forecasting models. Since the ground truth information is not available, the study further proposes a method to perform unsupervised ensembling to combine causal graphs discovered from existing methods. We believe our research will not only contribute to data-driven study of the Arctic but will also promote local communities and the broader international society to respond and adapt to a changing Arctic, thus having far-reaching impacts on forecasting global food and water demands, and economic security across the globe.

Spatiotemporal Data Mining using Multi-View Learning on Climate Change Studies

Data is generated at a phenomenal rate today from different sources such as sensors, global positioning systems, remote sensing devices, social media, etc. Spatiotemporal data in particular possesses unique characteristics such as spatial autocorrelation, spatiotemporal heterogeneity, temporal non-stationarity, and data residing on multiple scales and resolutions. Moreover, spatiotemporal data can be derived from multiple sources or views. Traditional data mining methods rely on analyzing and forecasting features from a single view alone. However, considering the mutually enriched information from multiple sources can lead to developing a broad understanding surrounding the spatial object which in turn can lead to making accurate spatiotemporal predictions. We address three important aspects in this study: (a) using spatiotemporal data from polar regions, we integrate all the views – spatial and temporal to determine anomalies/discords; (b) create neighborhoods to preserve the spatial correlation and heterogeneity present in the data; and (c) use neighborhood definition to analyze change along with multi view learning.
Detecting anomalous patterns from multiple and dynamic data sources adds to the complexity of making effective conclusions. In this study, we propose to build a novel framework to make spatiotemporal analysis related to the polar regions.  First, we generate neighborhoods to tap into spatiotemporal autocorrelation and heterogeneity. Second, we apply a multi-view learning model to take advantage of data coming from multiple related sources and perform multi-domain spatiotemporal analysis. Third, we detect anomalies that would aid in understanding which areas in the polar regions show unusual behavior and also the time intervals when these anomalies are present for a large spatial extent, i.e. Greenland wide. We use two scenarios where we studied the Arctic and Antarctic regions to study variables related to snow melt and sea ice thickness respectively to show a proof of concept of how multi-view learning can help understand the processes at play in understanding anomalous behavior in the cryosphere. Our initial results  in both Arctic and the Antarctic regions evaluating anomalous patterns present across regions and seasons corroborate findings of the polar science domain experts. We are interested in specifically understanding the role played by the climate in influencing the cryosphere variables using our proposed framework, incorporating multiple views coming from multiple domains in a given region.

Dr. Vandana Janeja, Chair/Advisor, Dr. Bayu Adhi Tama, Dr. Jianwu Wang, Dr. Karuna Joshi, Dr. Nicole Jeanne Schlegel

Date: Wednesday, October 26

Time: 10:00 a.m. - 11:00 a.m. PDT

Registration: https://hai.stanford.edu/events/thomas-teisberg-monitoring-ice-sheets-uav-borne-ice-penetrating-radar-systems-and-physics

Abstract:

The Antarctic Ice Sheet will play a growing role in sea level rise over the next century, but models of sea level contributions from the vast ice sheet carry far larger uncertainty estimates than other major contributing sources. A number of factors contribute to this uncertainty, all of which can be traced back to a sparsity of data. The Antarctic Ice Sheet is nearly 50% larger than the United States in area and holds ice equivalent to over 60 meters of global sea level rise. Satellite observations have turned the surfaces of Earth’s ice sheets into data-rich environments, yet the subsurface environments remain sparsely observed.

Multiple potential positive-feedback processes have been proposed that could dramatically alter our predictions of the future of the ice sheet. These hypotheses are difficult to test because the processes would likely not have occurred in the relatively short direct span of direct observations available to us, but, in many cases, we also lack the observational infrastructure to identify these processes beginning today. Even well-established processes still carry large uncertainties in how they will impact the ice sheet due to the poor spatial and temporal resolution of data available.

Our group explores this problem from two fronts: using robotics to enable more widespread data collection and leveraging data-based approaches to maximize the value of data collected. Uncrewed aerial vehicles (UAVs) carrying ice-penetrating radar instruments hold the potential to dramatically expand sub-surface data collection by reducing the cost, logistical complexity, and safety risks associated with current approaches. At the same time, the scale of the problem is so vast that it is critical to consider how we can optimize the deployment of resources to maximize the value of the collected data. This requires a move towards data-driven approaches to understanding the behavior of the ice sheet