Disentangling Land‐Atmosphere Interactions via Multilevel Temporal Information Partitioning Networks

Abstract

Understanding Earth’s hydrometeorological system requires characterizing the interactions at the land-atmosphere interface. However, the multivariate nature of land-atmosphere interactions across temporal scales and conditions remains insufficiently explored. This study proposes a multilevel framework integrating wavelet decomposition with Temporal Information Partitioning Networks (TIPNets) to investigate interactions among key atmospheric, soil, and vegetation variables using 5-year hourly observations across the United States. These variables are decomposed into high-, medium-, and low-frequency components for multilevel TIPNets analyses, assessing their unique, synergistic, and redundant contributions to land-atmosphere interactions across short-, medium-, and long-term levels. Results vary across temporal scales: at the short-term level, the multivariate system is generally dominated by solar radiation (SR), with synergistic interactions among SR, relative humidity (RH), and evapotranspiration playing an important role. Meanwhile, the interaction patterns shift as precipitation increases. Medium-term interactions exhibit redundancy, influenced primarily by RH and surface soil moisture (SM). Land cover influences are notable at this level, with contributions from surface SM becoming vital in croplands while deep SM being important in forests. At the long-term level, vegetation and deep soil temperature emerge as dominant sources, with redundant contributions reflecting strong vegetation-soil thermal coupling. Furthermore, dry conditions amplify land-atmosphere feedbacks and enhance system connectivity. This study provides a quantitative, multilevel characterization of dynamic land-atmosphere interactions from an information-theoretic perspective, offering practical insights for enhancing hydrological modeling and climate adaptation strategies.

Source link