Using Temporal Information Partitioning Networks (TIPnets) to Assess Land–Atmosphere Coupling

The terrestrial leg of land atmosphere coupling, determined by the dependency between land conditions and surface fluxes, is characterized by threshold behaviors and nonlinearity. A recent approach, Temporal Information Partitioning Networks (TIPNets), quantizing the reduction in uncertainty between source and target variables with a moving time window, is tailored for exploring such time-varying nonlinear dependency. The quantized value, called multivariate mutual information, can further be partitioned into unique, mutual, and redundant contributions of the source variables. TIPNets has been applied only locally; we apply TIPNets globally to understand spatiotemporal behavior of land-atmosphere interactions. For each season, TIPNets are constructed from different combination of three time-series variables by using daily MERRA-2 reanalysis data; candidate variables associated with land atmosphere coupling are soil moisture, surface temperature, radiation, wind speed, leaf area index, and surface heat fluxes. Preliminary analysis, using surface net radiation and soil moisture as the information sources and their coupling to variations in latent heat flux as the target, shows strong dependencies over high latitudes in boreal spring and semi-arid regions during boreal summer. Partitioning of multivariate mutual information indicates that the redundant component dominates the dependencies. Analyses for different combinations of variables are ongoing to assess comprehensive behavior of the land atmosphere interactions.