Causal inference of root zone soil moisture performance in drought

Abstract

Soil moisture plays a crucial role in surface hydrological processes and land–atmosphere interactions. It can influence vegetation growth directly, serving as a significant indicator for monitoring agricultural drought. However, spatially continuous datasets of root zone soil moisture rely on model simulations, introducing numerous uncertainties associated with model parameters and input data. Currently, multiple soil moisture products derived from model simulations exist, but their representation at spatial scales remains unclear. Moreover, their abilities to express soil–atmosphere and soil–vegetation interactions within land–atmosphere coupling are not understood, leading to divergent inclinations toward drought. This study investigates the performance of five soil moisture products, European Centre for Medium-Range Weather Forecasts Reanalysis v5-Land (ERA5-Land), Global Land Data Assimilation System (GLDAS), Global Land Evaporation Amsterdam Model (GLEAM), The Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2), and SoMo.ml, under drought conditions. The bias, correlation, and difference of standard deviation (STDD) were calculated between these products and the observations from International Soil Moisture Network stations. The causal probability of soil, meteorological, and agricultural drought was calculated using the causal-effect Peter and Clark (PC) Momentary Conditional Independence (MCI) method to evaluate the data propensity of these products. ERA5-Land and SoMo.ml gave a similar performance with the highest accuracy, which was attributed to the use of the same meteorological forcing data. The biases of soil moisture from these two products at surface, middle and deep depths against station observations are below 0.1 m³/m³, and the STDD is within 0.05 m³/m³. The accuracy of GLDAS is comparatively lower, characterized by lower correlations (below 0.2 for deeper layers) and high bias (above 0.15 and 0.2 for middle and deep layers, respectively). This discrepancy could be attributed to substantial biases in the precipitation forcing data. ERA5-Land shows higher spatial resolution and greater spatial heterogeneity, whereas MERRA-2 underperformed in this area. MERRA-2 had the strongest connection to agricultural drought, with a propensity probability of 0.477. Conversely, SoMo.ml demonstrates the strongest connection to meteorological drought, with a propensity probability of 0.234. Due to the errors in simulated and observational data during the MERRA data assimilation, substantial biases in the soil moisture data, and low accuracy in meteorological forcing of GLDAS, there was no clear causal relationship between soil moisture drought and meteorological drought between these two products. These findings provide recommendations for the use of soil moisture products in drought research.