What Limits the Accuracy of Widely Used Soil Moisture Products? An LSTM Diagnostic of SMAP and ERA5-Land at COSMOS-UK

Authors:
Preprint
Publisher website
Google Scholar
Abstract
Soil water content underpins hydrological, ecosystem, and land–atmosphere processes, and estimates based on SMAP and ERA5-Land support climate modelling and land management. Yet these products often disagree with in situ observations in magnitude and dynamic range, and it remains unclear whether this reflects poor representation of wetting–drying dynamics or the absolute water-content range (or both). A Long Short-Term Memory (LSTM) model was developed to represent relative wetting–drying dynamics from horizontally polarised SMAP brightness temperature and ERA5-Land precipitation and potential evaporation. LSTM-model predictions were converted into absolute soil water content using site-specific minimum and maximum water contents (hydrological bounds) deduced from observations, SMAP soil moisture retrieval (SMAP-SM), or ERA5-Land soil moisture predictions (ERA5-SM). The model was trained at three COSMOS-UK sites and evaluated at 16 independent COSMOS-UK sites using field-scale cosmic-ray neutron sensing (CRNS) and point-scale time-domain transmissometry (TDT), with SMAP-SM and ERA5-SM as benchmarks. Despite the small number of input variables, the LSTM-model was equally good or better than the other products. Using observed bounds, the LSTM-model achieved median RMSEs of 4.4 (CRNS) and 6.7 percentage points (TDT), with near-zero bias; SMAP-SM was dry-biased and ERA5-SM wet-biased. Product-derived bounds increased absolute errors. With respect to anomaly dynamics, ERA5-SM slightly outperformed the LSTM-model, while SMAP-SM was weakest. The results indicate that SMAP-SM and ERA5-SM track relative dynamics more reliably than absolute soil water content. Future research must invest in both (i) estimating relative soil moisture dynamics and (ii) assessing the hydrological bounds to translate those dynamics into absolute values.

This publication has 37 references indexed in Scilit: