Temporal Networks: A New Approach to Model Non-stationary Hydroclimatic Processes with a Demonstration for Soil Moisture Prediction

Interactions between different components of the hydrologic cycle show a time-varying characteristic due to the impact of climate change that lead to the non-stationarity in many hydroclimatic variables. In fact, a lack of stationarity in most of the hydroclimatic processes is realized in many cases. In such situation, alternative methodologies that can effectively learn (adapt) from the changing climate will help in development of effective and efficient hydroclimatic models. This study presents the potential of a recently developed approach, namely temporal networks. These time-varying network structures help in hydroclimatic modelling by (i) identifying the complex association (dependence structure) among the large pool of influencing variables and (ii) identifying the temporal variability of the dependence structure to capture the time-varying characteristics in the association among the hydroclimatic variables. The approach helps to improve the accuracy of the model performance under a changing climate. As a demonstration, we picked out the slowly changing soil moisture regime at a location and attempted to capture its time-varying characteristics through temporal networks based time-varying modelling framework. Our target is to predict the monthly soil moisture with one-month to one-season (three months) in advance. The performance of the temporal networks based model is contrasted with the time-invariant modelling philosophy. Towards this, (i) time-invariant network model, as the closest counterpart, and (ii) Support Vector Regression (SVR) based models, Machine Learning (ML) technique commonly implemented in the field of hydroclimatology, are used. We established that the temporal networks satisfactorily capture the soil moisture variability over time.