guest ::
| Home > Publications database > Towards improved representation of the hydrological cycle by assimilating multi-source remotely sensed soil moisture data in terrestrial system models |
| Home > Publications database > Towards improved representation of the hydrological cycle by assimilating multi-source remotely sensed soil moisture data in terrestrial system models |
| Book/Dissertation / PhD Thesis | FZJ-2026-02772 |
2026
Forschungszentrum Jülich GmbH Zentralbibliothek, Verlag
Jülich
ISBN: 978-3-95806-953--4
This record in other databases:
Please use a persistent id in citations: doi:10.34734/FZJ-2026-02772
Abstract: Soil moisture is a key variable in the water and energy cycles at the atmosphere-land-vegetation interface. Accurate and continuous soil moisture data are essential for many applications, including water resources management, drought monitoring, and agricultural forecasting. In situ soil moisture measurements from laboratories or sensors are limited by spatial coverage and high cost. Land surface models (LSMs) can simulate soil moisture from point to large scales, but are subject to simplified parameterizations and uncertainties due to model structure and input data. For example, the role of lateral water flow is typically not represented in LSMs. The advent of remote sensing technologies has made it possible to obtain soil moisture information on a large scale and in near real-time. However, soil moisture retrievals are coarse and limited by dense vegetation, complex topography, and frozen soils. Data assimilation (DA) can use data from sources like remote sensing to improve the performance of land surface models. We evaluated LSM simulations over a temperate-climate region with complex topography and land forms in western Germany, using a variety of data types, e.g., soil moisture and evapotranspiration (ET). One objective of this PhD research was to assess the influence of considering lateral flow and the subsurface on the simulation of soil moisture and other hydrological variables, and to investigate whether the coupled land surface-subsurface model could extract more remotely sensed information in the DA process. Another objective was to gain insight into the value added by merging remotely sensed soil moisture information using different merging schemes, and to assess whether merged soil moisture datasets allow for improving land surface characterisation more than the assimilation of single remotely sensed soil moisture products. In the first study, a cross-evaluation of land surface model results (with and without lateral flow processes), the National Aeronautics and Space Administration (NASA) Soil Moisture Active/Passive (SMAP) mission soil moisture product, and cosmic-ray neutron sensor (CRNS) measurements was carried out. A traditional land surface model (the Community Land Model (CLM) version 3.5) and a coupled land surface-subsurface model (CLM-ParFlow) were applied. Compared to CLM stand-alone simulations, the coupled CLM-ParFlow model considers both vertical and lateral water movement. In addition to standard validation metrics, a triple collocation (TC) analysis has been performed to help understand the random error variances of different soil moisture datasets. In this study, it was found that the three soil moisture datasets are consistent. The coupled and uncoupled model simulations were evaluated at CRNS sites, and the coupled model simulations showed less bias than the CLM stand-alone model (-0.02 cm3/cm3 vs 0.07 cm3/cm3), similar random errors, but a slightly smaller correlation with the measurements (0.67 vs 0.71). The TC analysis showed that CLM-ParFlow reproduced better soil moisture dynamics than CLM standalone and with a higher signal-to-noise ratio. This suggests that the representation of subsurface physics is of significant importance in land surface modelingand that coupled land surface-subsurface modeling is of high interest. Land surface modelling combined with data assimilation provides the potential to generate highly accurate soil moisture estimates across regional and global scales. In a second study in this PhD dissertation, the CLM and the coupled land surface subsurface model CLM-ParFlow, which considers lateral surface and subsurface flows, were applied on the study region. The soil moisture retrievals from the Soil Moisture Active/Passive (SMAP) mission were assimilated with the Localized Ensemble Kalman Filter (LEnKF) (with and without parameter estimation) using an ensemble size of 32 realizations. The simulated soil moisture, ET, and groundwater level were evaluated using in situ observations from a CRNS network, Eddy Covariance (EC) stations, andgroundwater measurement wells. The results showed that the assimilation of remotely sensed soil moisture product improved the correlation from 0.71-0.78 to 0.79-0.82 and decreased the unbiased Root Mean Square Error (ubRMSE) from 0.062-0.048 cm3/cm3 to 0.058-0.045 cm3/cm3. The characterisation of ET showed a limited improvement with the highest ubRMSE reduction of 5%. The assimilation did not improve the groundwater level characterization. In addition, the joint state-parameter update did not give a better performance than updating the states alone, suggesting that SMAP retrievals do not provide sufficient information to update the parameters. Microwave remote sensing technology has emerged to provide valuable products to monitor and assess soil moisture content at regional or global scales. However, each soil moisture product exhibits different advantages and shortcomings. Data fusion could help improve accuracy by merging information from different sources. In a third study in this PhD thesis, a traditional TC-based method and a novel Long Short Term Memory network (LSTM) were used to merge soil moisture products from the SMAP mission, Advanced Microwave Scanning Radiometer 2 (AMSR2) and The Advanced SCATterometer (ASCAT). This research reveals that the LSTM outperforms the traditional TC-based method for data fusion when evaluated against field measurements. The study identifies that both climate forcing and physiographic attributes significantly influence the spatial and temporal variations observed in the LSTM prediction scheme. Consequently, the study highlights the considerable potential of the LSTM method for large-scale integration of remote sensing soil moisture data. The comparative efficacy of assimilating microwave retrievals from different missions remains unclear. A fourth study in this dissertation investigated the effectiveness of assimilating soil moisture retrievals from both active and passive microwave instruments and their merged soil moisture datasets into the CLM using an ensemble Kalman filter (EnKF) approach. Insitu observations from the TERrestrial EnviroNmental Observations (TERENO) network were employed to evaluate the modelled soil moisture by different sets of experiments: (1) individual assimilation of each product and (2) assimilation of combined active and passive retrievals by arithmetic mean, TC-based method, and LSTM scheme. Results showed that the assimilation generally improved the model performance in terms of Pearson r and ubRMSE. The SMAP and ASCAT products were more informative than the AMSR product, with a median improved Pearson r of 0.08 and 0.09 and a reduction in ubRMSE of 28.5% and 27.3% compared to open loop runs, respectively. The similar performance for SMAP and ASCAT demonstrated that they may be complementary tools. The assimilation of merged datasets gave a similar performance to SMAP and ASCAT retrieval and a slightly better performance than ASCAT retrieval. This dissertation aims to investigate the use of coupled land surface subsurface models to enhance the performance of LSMs through DA. However, the results indicate that the improvements for other compartments are limited, and many challenges remain. For the complex LSMs, a multivariate and multi-scale approach needs to be explored, which considers the integration of information from multiple observation types (e.g., soil moisture, leaf area index, total water storage) and their spatial mismatch with the model grid. Furthermore, it is observed that assimilating a merged product of microwave soil moisture retrievals hardly gives better results than assimilating single retrievals. In the DA scheme, a uniform observation error was applied, which might be unrealistic. A temporally and spatially variable observation error could be used to help DA assign higher weights to more accurate information and lower weights to unreliable information. Meantime, a more accurate remote sensing product (Sentinel-1 at 1 km) might be preferable for constraining the high-resolution model.
; 
|
The record appears in these collections: |
PDF