Anomaly Detection Using Graph Deviation Networks Within Spatiotemporal Neighborhoods: A Case Study in Greenland

Polar ice melt contributes to sea level rise. To understand this contribution, we need to examine the anomalous behaviors leading to significant snowmelts in polar regions, including the Greenland ice sheet. These regions are complex systems where various phenomena are represented by different sets of spatiotemporal data. Such data possess unique characteristics like spatial autocorrelation, heterogeneity, temporal nonstationarity, and multiple scales and resolutions. In this article, we provide a framework to analyze disparate datasets by forming spatial neighborhoods to capture local behaviors. We then perform graph deviation network-based anomaly detection for multivariate datasets within these neighborhoods. Although this study focuses on spatiotemporal data from Greenland as an example, the methodology is intended to be adaptable and relevant to other regions with similar data properties. Specifically, using spatiotemporal data from Greenland, we, first, integrate all subdomain data, including both spatial and temporal data. Second, create neighborhoods to preserve the spatial autocorrelation and heterogeneity present in the data. Third, apply graph deviation networks, a variant of graph neural networks, to locate anomalous regions with respect to snowmelt. We outline our findings in the Greenland region, evaluating anomalous patterns and validating them with ground truth findings from polar science domain experts. Our methodology allows for performing localized analysis on a Greenland-wide scale.