Temporal information partitioning networks to infer ecohydrologic behaviors

An ecohydrologic system is a complex network, in which the shifting behavior of individual components and the connectivity between them determines the dynamics. This connectivity between components can act to constrain, accentuate, or otherwise modify the variability of individuals. In an ecohydrologic system, connectivity exists in the form of many time dependent relationships between states and fluxes related to water, energy, nutrients, soils, and vegetation. Although relationships are constrained by conservation laws, they exhibit a wide range of variability at many timescales due to non-linear interactions, threshold behavior, forcing, and feedback. Moreover, these aspects of connectivity and variability exist at a single location or over a spatial gradient. The understanding of this connectivity within the system as a whole necessitates an appropriate framework, in which evolving interactions are identified from time-series observations. The goals of this thesis are to (i) develop a Temporal Information Partitioning Network (TIPNet) framework for understanding the joint variability of network components as characterized by time-series data, and (ii) apply this framework to understand ecohydrologic systems across climate gradients based on flux tower and weather station observations. In the TIPNet framework, nodes in the network are time-series variables, and links are information theoretic measures that quantify multivariate lagged time-dependencies from lagged "source" nodes to "target" nodes. The strength of this framework is its ability to characterize information flow between variables over short time windows, and further distinguish aspects of unique, redundant, and synergistic dependencies. Redundant information is overlapping information provided by multiple sources to a target, unique information is only provided by a single target, and synergistic information is provided only when two or more sources are known together. Based on data from three Critical Zone Observatories, we find that network structure shifts according to conditions at sub-daily time scales and constraints imposed by seasonal energy and water availability. TIPNets constructed from 1-minute weather station data reveal shifts in time-scales and levels of uniqueness, synergy, and redundancy between wet and dry conditions. A more complex network of synergistic interactions characterizes several-hour windows when surfaces are wet, and peaks in information flow during the growing season correspond to shifts in precipitation patterns. Networks based on half hourly flux tower data reveal seasonal shifts in the nature of forcing to carbon and heat fluxes from radiation, atmospheric, and soil subsystems. Along two study transects, we attribute variability in heat and carbon fluxes within constraints imposed by energy and moisture availability to joint interactions that are more synergistic in the spring and redundant in the fall. Finally, we explore the nature of information flow along an elevation gradient from flux towers located along a transect to gauge local versus non-local connectivity. While the strength of shared information between variables at a site reflects local connectivity, shared information between variables at different sites reflects non-local connectivity. Along two elevation transects, we find that information flow between distant sites indicates directional connectivity that is related to dominant weather patterns. At the Southern Sierra CZO in California, non-local information flow is dominantly west to east, corresponding to weather forcing from the Pacific Ocean eastward, while non-local flow has less directionality at Reynolds Creek CZO, where sites are much closer together and there is no dominant weather forcing direction along the transect. The developed framework and applications presented in this thesis reveal the common presence of multivariate process interactions at timescales from minutes to hours, many of which would not be detected using traditional approaches. For an ecohydrologic system, the complex network of relationships dictates ecosystem resilience to perturbations such as climate change, drought, or human influences. More broadly, the methods and framework developed here contribute toward a holistic understanding of complex systems, and are applicable to a range of studies of evolving networks