Complex Networks Reveal Persistent Global / Regional Structure and Predictive Information Content in Climate Data

Recent articles have posited that the skills of climate model projections, particularly for variables and scales of interest to decision makers, may need to be significantly improved. Here we hypothesize that there is information content in variables that are projected more reliably, for example, sea surface temperatures, which is relevant for improving predictions of other variables at scales which may be more crucial, for example, regional land temperature and precipitation anomalies. While this hypothesis may be partially supported based on conceptual understanding, a key question to explore is whether the relevant information content can be meaningfully extracted from observations and model simulations. Here we use climate reconstructions from reanalysis datasets to examine the question in detail. Our tool of choice is complex networks, which have provided useful insights in the context of descriptive analysis and change detection for climate in the recent literature. We describe a new adaptation of complex networks based on computational approaches which provide additional descriptive insights at both global and regional scales, specifically sea surface variables, and provide a unified framework for data-guided predictive modeling, specifically for regional temperature and precipitation over land. Complex networks were constructed from historical data to study the properties of the global climate system and characterize behavior at the global scale. Clusters based on community detection, which leverage the network distance, were used to identify regional structures. Persistence and stability of these features over time were evaluated. Predictive information content of ocean indicators with respect to land climate was extracted using a suite of regression models and validated on held-out data. Our results suggest that the new adaptation of complex networks may be well-suited to provide a unified framework for exploring climate teleconnections or long-range spatial and temporal dependencies of ocean and land variables, producing descriptive analysis leading to potentially new insights in climate science, as well as generating data-guided predictions which may leverage information content in climate model simulations themselves to improve predictions of variables that are crucial for impacts assessments but less well predicted from models. However, the applicability of this tool must be evaluated on a case by case basis, and any future hypotheses about descriptive or predictive insights needs to be carefully tested and validated with climate science interpretations.