Assimilation of SMOS soil moisture over the Great Lakes basin

Xiaoyong Xu, Bryan A. Tolson, Jonathan Li, Ralf M. Staebler, Frank Seglenieks, Amin Haghnegahdar, Bruce DavisonRemote 2015 · DOI

TL;DR

This study assimilates SMOS satellite soil moisture data into a land surface model (MESH) over the Great Lakes basin using an Ensemble Kalman Filter. The research highlights how vegetation type significantly impacts SMOS retrieval accuracy and demonstrates that assimilation generally improves model soil moisture estimates, especially in crop-dominated areas.

What Problem Does This Paper Address?

Accurate soil moisture information is crucial for weather and climate forecasting. Satellite remote sensing, particularly with the L-band SMOS mission, offers a promising way to obtain large-scale soil moisture data. However, effectively integrating these satellite observations into land surface models remains a challenge. Key issues include:

  • Scale Mismatch: Satellite footprints are much coarser than point-based in-situ measurements, making direct validation difficult.
  • Vegetation Effects: Dense vegetation can significantly attenuate L-band microwave signals, affecting SMOS retrieval accuracy.
  • Model-Satellite Bias: Differences in how models and satellites represent soil moisture can lead to biases that hinder effective assimilation.
  • Understanding Assimilation Performance: It's important to understand how well assimilation improves model estimates and how this improvement varies with different land cover types and the quality of both satellite observations and model forecasts.

This paper aims to address these challenges by assimilating SMOS data into the MESH model over the Great Lakes basin, a region with diverse vegetation, and by rigorously evaluating the assimilation's performance across different land cover types.

Methods

The core of this study involves data assimilation, specifically using the one-dimensional Ensemble Kalman Filter (1D-EnKF). Here's a breakdown:

  • 1D-EnKF: This is a sequential data assimilation technique that uses an ensemble of model states to estimate the model's error statistics. It's well-suited for large-scale applications where a full Kalman filter would be computationally prohibitive. The "1D" aspect means assimilation is performed independently for each model grid cell.
  • Bias Correction (a priori rescaling): Before assimilation, SMOS soil moisture retrievals are rescaled to match the cumulative distribution function (CDF) of the model's surface soil moisture. This step aims to reduce systematic biases between the satellite data and the model, ensuring that the assimilation focuses on improving the temporal dynamics rather than forcing the model to match potentially biased absolute values.
  • Land Surface Model (MESH): The Modélisation Environmentale-Surface et Hydrologie (MESH) model is used. It's a distributed hydrological model that accounts for subgrid-scale variability using a Grouped Response Unit (GRU) approach, where each grid cell is composed of different land cover types.
  • Validation: Soil moisture estimates from SMOS, the open-loop model (model run without assimilation), and the assimilation product are validated against point-scale in-situ measurements using the correlation coefficient (skill R). This metric focuses on the temporal agreement between observations and model estimates.

Datasets

  • SMOS Level 2 Soil Moisture User Data Product (MIR_SMUDP2): This product provides instantaneous soil moisture retrievals from the European Space Agency's Soil Moisture and Ocean Salinity (SMOS) satellite. Data from 2010-2013 were used.
  • In-situ Soil Moisture Measurements: Data from three networks were used for validation:
  • Michigan Automated Weather Network (MAWN)
    • Soil Climate Analysis Network (SCAN)
    • Fluxnet-Canada Research Network (FCRN)
  • Meteorological Forcing Data: Hourly precipitation data from the Canadian Precipitation Analysis (CaPA) and other meteorological variables from the Global Environmental Multiscale (GEM) model were used to drive the MESH model.

Key Insights and Ideas

  • Vegetation is Key: The study strongly confirms that vegetation cover significantly impacts SMOS soil moisture retrieval accuracy. Crop surfaces generally yield higher SMOS skill (correlation with in-situ data) than forest surfaces, likely due to L-band microwaves penetrating vegetation better in less dense canopies.
  • Assimilation Benefits Vary: Data assimilation generally improves model soil moisture estimates. However, the degree of improvement (skill gain) is not uniform. It's most pronounced in crop-dominated areas where SMOS observations are more reliable and the open-loop model might be less accurate.
  • Open-Loop Model Skill Matters: The effectiveness of assimilation is inversely related to the skill of the open-loop model. When the model is already performing well, the potential for improvement through assimilation is reduced. Conversely, when the model's open-loop performance is poor, assimilation can offer substantial gains, provided the satellite observations are sufficiently accurate.
  • Surface vs. Root Zone: Assimilating surface soil moisture data from SMOS effectively propagates information to the root zone, improving estimates there as well.
  • Bias Correction is Crucial: Rescaling SMOS retrievals to match the model's CDF is essential. Without this, biases could lead to suboptimal assimilation results. This also means that the absolute magnitudes of the assimilated product are less meaningful than their temporal variations.

Key Results

  • SMOS Skill: SMOS soil moisture retrievals showed moderate to high skill (median R of 0.55-0.64) over crop-dominated grids (sCmC, sCmF) but significantly lower skill (median R of 0.23-0.32) over forest-dominated grids (sFmC, sFmF).
  • Open-Loop Model Skill: The open-loop MESH model generally performed better over forest-dominated grids (median R of ~0.60-0.62) than over crop-dominated grids (median R of ~0.39-0.40) for surface soil moisture.
  • Assimilation Improvement (ARA-M): The assimilation product consistently showed improved skill over the open-loop model for both surface and root-zone soil moisture. The largest improvements were observed in crop-dominated grids (sCmC), where the assimilation skill could reach median R values of ~0.64. Forest-dominated grids showed weaker or even negative skill gains.
  • Assimilation vs. SMOS (ARA-S): The assimilation product's skill was generally better than the SMOS observation skill alone. However, the skill improvement relative to SMOS (ARA-S) was larger in forest-dominated grids where the open-loop model was strong, and smaller in crop-dominated grids where SMOS skill was already high.

Limitations and Future Work

  • Point-Scale Validation: While the study uses point-scale in-situ data, it acknowledges that this can lead to a spatial mismatch with the coarser satellite and model grid cells. The authors argue that temporal variability is spatially representative, and the use of correlation coefficient mitigates some of these issues. However, denser in-situ networks would be advantageous.
  • Bias Correction: The study uses a priori bias correction. While effective, it means the absolute magnitude of the assimilated soil moisture is less reliable than its temporal dynamics. Future work could explore more sophisticated bias-aware assimilation systems.
  • Ensemble Size: The 1D-EnKF used 12 ensemble members. While standard, a larger ensemble size could potentially improve the accuracy of error estimates and reduce the occurrence of negative skill improvements.
  • Forcing Data Quality: The study notes that differences in open-loop model skill between 2010/2011 and 2012/2013 might be partly due to variations in meteorological forcing data quality. Further investigation into the impact of forcing data is warranted.
  • Future Directions: The authors suggest that the ability of SMOS to measure soil moisture across various vegetation types is valuable for understanding assimilation dynamics. Future work could explore more advanced assimilation techniques and investigate the impact of SMOS data on weather and climate forecasting initializations.

Summary updated

BibTeX
@inproceedings{xu2015assimilation,
  author = {Xiaoyong Xu and Bryan A. Tolson and Jonathan Li and Ralf M. Staebler and Frank Seglenieks and Amin Haghnegahdar and Bruce Davison},
  title = {{Assimilation of SMOS soil moisture over the Great Lakes basin}},
  booktitle = {Remote},
  year = {2015},
  doi = {10.1016/j.rse.2015.08.017},
}
Figure