Organisation

November

Soil moisture remote sensing using active microwaves and land surface modeling

Summary and conclusions of Rogier van der Velde's PhD thesis

Soil moisture has an important effect on the partitioning of solar radiation and influences, thus, the development of weather systems. It is, therefore, expected that an improved representation of the soil moisture dynamics in atmospheric circulation models (ACM’s) will enhance their predictive skills. There are different methods to achieve this. For example, the reliability of simulated soil moisture can be improved by using more realistic model structures. Increasingly popular with the availability of more remote sensing observations is, however, to reduce the uncertainties of simulations through integration with measurements (or satellite retrievals).

This thesis contributes to both aspects. Chapters 5 to 8 deal with the retrieval of soil moisture from active microwaves, whereas Chapter 9 and 10 discuss land surface model simulations performed by the Noah model. In Chapter 11, the soil moisture retrievals are compared to simulations by the Noah model coupled to the MM5 regional climate model. These parts are summarized below.

Soil moisture retrieval from active microwaves

A ground-based C- and L-band scatterometer data set collected throughout the corn growth cycle and a set including 2.5 years of ASAR (C-band and VV polarization) acquisitions over the central part of the Tibetan Plateau have been used for studying the soil moisture retrieval. The ground-based scatterometer data set benefits from a comprehensive set of in-situ measurement, specifically vegetation biomass and soil moisture. Unique about the ASAR data set over the data scarce Tibetan Plateau is that the high resolution microwave measurements are acquired at a fairly high temporal and over a substantial period of time. The ground-based scatterometer data has been utilized for evaluating the vegetation and surface roughness effects on backscatter (σo) measurements. Over a larger temporal and spatial scale, the impact of different land surface conditions (e.g. soil moisture, sparse vegetation, freeze/thaw) on the σo has been investigated and the retrieval of soil moisture has been performed using the ASAR data set.

Chapter 5 discusses the effect of vegetation on fully polarimetric (HH, HV, VV) C- and L-band σo measured at incidence angles of 15, 35, and 55 degrees for the corn growth cycle. The analysis shows that depending on the antenna configuration and growth stage, the σo measurements can be dominated either by attenuated surface scattering or by scattering from vegetation. The first mechanism is strongest within C-band σo measurements collected at 15 degrees during the early growth stage. The latter is more notable among the σo measured at peak biomass and incidence angles of 35 and 55 degrees. Somewhat surprising is, however, that even at peak biomass the measured σo response to soil moisture is still considerable enabling the retrieval of soil moisture. This σo sensitivity to soil moisture is ascribed to scattering along the soil-vegetation pathways.

Based on these observations an alternate method is proposed to correct σo measurements for vegetation and obtain more reliable soil moisture retrievals. The method is based on the concept that the ratio of the surface scattering contribution over the observed σo is affected by vegetation and can be described as a function of the vegetation water content (W). Experimentally determined relationships between this ratio and W have been used to reconstruct the surface scattering component from the σo measurements and retrieve soil moisture. Validation of the retrievals obtained with this method against the measured soil moisture yields errors varying from 0.033 to 0.063 m3 m-3 depending on the antenna configuration. This accuracy is quite good specifically given the dense vegetation coverage with at peak biomass with a W of 5.1 kg m-2. Additional studies are, however, needed to establish the validity of the proposed method for other crop types.

Also, the roughness of the soil surface influences the observed σo and has to be considered when retrieving soil moisture. The (Advanced) Integral Equation Method ((A)IEM, Fung et al. 1992, Chen et al. 2003) is the most widely applicable surface scattering model and is often used to simulate the σo from bare soil surfaces. This model represents the complex geometry of the soil surface as a single-scale stationary process, which allows adopting a parameterization based on 1-D surface height profile consisting of the root mean square height (s), correlation length (l) and autocorrelation length function (ACF). However, over large spatial domains typically insufficient information is available to determine this parameterization reliably. For retrieving soil moisture, therefore, a reduced number of parameters has to be used and the surface roughness is often assumed to be temporally stable.

Chapter 6 discusses the impact of these simplifications on the accuracy of the retrieved soil moisture using the vegetation corrected σo measurements. The surface roughness parameters have been inverted for four parameterization types by assuming an ‘Exponential’ and ‘Gaussian’ ACF’s, and by using fixed (measured) or variable l. An evaluation of the retrieval accuracies shows that the most significant differences are noted when using different ACF’s, while the l has only a minor impact. However, differences in the retrieval accuracy are smaller than 0.01 m3 m-3. This suggests that regardless of the employed parameterization type, an effective value for the s can be found without having a large impact on the retrieval accuracy.

Additional analyses on the temporal stability of the roughness indicates that over the entire growth cycle surface roughness changes have only a limited impact on the retrieved soil moisture (<0.01 m3 m-3). On a daily basis, however, the largest differences between the measured and retrieved soil moisture occur specifically after rain events. This increase in the retrieval error is strongest at HH polarization, L-band, and large incidence angles, which is in agreement with previous reports on σo sensitivities to surface roughness (e.g. Holah et al. 2005, Beaudoin et al. 1990, Zribi et al 1997). In contradiction to previous reports (e.g. Ulaby and Batlivala 1976, Ulaby et al. 1996, Shi et al. 1997, Macelloni et al. 1999) the soil moisture retrieved throughout the corn growth cycle is, therefore, more accurate for the VV polarization.

Chapter 7 discusses the influence of land surface states on the ASAR σo at longer temporal and different spatial scales over the Tibetan Plateau. The σo signatures from 1x1 km2 areas covering a grassland and a wetland have been studied to identify its sensitivity to the changing land surface states. The lowest σo values from the grassland and wetland areas are obtained throughout the winter seasons because soil water is predominantly frozen resulting in dielectric properties comparable to dry soil conditions. Towards the summer, the wetland σo increases steadily and reaches its maximum as the monsoon is at peak intensity, while the grassland σo in the summer is characterized by large temporal variations. This contrast between the grassland and wetland σo dynamics is attributed to the highly variable soil moisture in the grassland caused by a large evaporative demand, while soil moisture conditions in wetlands are temporally stable.

The differences between the grassland and wetland σo dynamics has consequences for the spatial σo variability observed at different spatial scales (e.g. 1x1 km2, 5x5 km2 and 30x30 km2). The comparison of the mean σo with its standard deviation (stdev) results in a specific triangular data points distribution, whereby the peak located near the mid range of σo values. When these σo dynamics are considered to be representative for the soil moisture conditions, these results suggest that the relationship between mean soil moisture and the spatial variability is not always uniquely defined. This implicates that relationships between the mean soil moisture and stdev, used for downscaling coarse resolution soil moisture products, cannot be assumed to be time-invariant, but should be obtained from additional near real-time data sources, such as SAR data.

Another important observation is the discrepancy between the annual cycles of the NDVI and the wetland σo. The NDVI increase starts later than for the wetland σo and its decrease is observed earlier. This, supported by the small portion of σo variations explained by the NDVI for both the grassland and wetland suggests that the vegetation effects on the ASAR σo observed over the Tibetan Plateau is fairly small.

Therefore, the algorithm employed in Chapter 8 for retrieving soil moisture is solely based on the AIEM surface scattering model and assumes that the effects of vegetation are negligible. The roughness parameters needed for AIEM simulations is obtained through the inversion of a sequence of three σo measurements collected at different view angles under assumed dry conditions. As the dielectric properties of frozen soil are equivalent to the ones of dry soil, ASAR images collected in February and January have been utilized for the roughness inversion. The derived roughness parameterizations are used as input for retrieving soil moisture from the time series of ASAR σo.

The resulting soil moisture retrievals represent the monsoon sequence quite well with their maximum values registered in the months July to September. Spatially, the retrieved soil moisture dynamics is also in accordance with the expectations; with the soil moisture retrieved over grasslands being highly variable, and wet and fairly stable conditions observed over the wetlands. A comparison of the retrieved against the soil moisture measured at a wetland site and three grassland sites yields Root Mean Squared Differences (RMSD’s) of 0.060 and 0.032 m3 m-3 for the wetland site and the three grassland sites, respectively. Normalized for their different dynamic soil moisture ranges the uncertainty levels obtained for the wetland and grassland are, however, quite similar, 16.9 and 17.3 % respectively. These error levels are comparable to the results of previous SAR based soil moisture retrieval studies.

Simulation of land processes

At present, the Noah land model is coupled to the National Centers for Environmental Prediction (NCEP) operational weather and climate prediction model for quantifying the soil moisture impact on land-atmosphere interactions. Crucial for its performance is a proper simulation of water movement through the soil column. Chapter 9 evaluates different numerical schemes for vertically integrating the soil water flow and analyses the impact of the soil hydraulic model (SHM) employed for calculating the transport coefficients.

Noah simulations forced by atmospheric variables measured at the Cabauw meteorological station (The Netherlands) shows that the current method used for vertical integration of the soil water flow systematically underestimates the upward water transport; often referred to as capillary rise. Therefore, a rapid dry-down of the root zone is observed, while the deep soil layers remain relatively wet even under dry conditions and a high evaporative demand. Two alternative schemes are presented that incorporate the capillary rise mechanism.

The SHM shapes the soil hydraulic functions (SHF’s), which determine the magnitude of the transport coefficients and the impact of the soil moisture stress on the heat flux partitioning. Noah (and most other LSMs) uses the SHM by Campbell (1974), while it is commonly understood that the Van Genuchten (1980) SHM represents measured SHF’s better. Surface water and energy budgets simulated by Noah using these two SHM’s are compared. The difference in the parameterization utilized by the Van Genuchten and Campbell SHM prevents, however, a direct comparison. The SHF’s of five soil type provided by the Dutch pedotransfer function (Wösten et al. 2001) have, therefore, been used to derive the Van Genuchten and two type Campbell parameterizations, whereby, 1. The theoretical similarity in the retention curves is used (Campbell A); 2. The Campbell parameters are fitted to the relationships between the transport coefficient and soil moisture content (Campbell B).

Compared to Van Genuchten, typical for Campbell A are the larger transport coefficients in the mid and dry soil moisture range, while for Campbell B these coefficients are lower near saturation due to the necessity of smaller saturated conductivity. A consequence of these differences is that Noah with Campbell A transports water faster through the soil column, while with Campbell B it generates more surface runoff. Noah simulates, thus, larger soil water losses (e.g. evaporation, drainage and surface runoff) using both forms of Campbell SHF’s. This as well as the underestimation of the capillary rise mechanism could be one of the explanations for the prediction of too warm and too dry summers by ACM’s.

Apart from issues related to the model physics, various investigations have also shown that the performance of land models can be improved by adjusting its parameterizations, specifically in the extreme environment, such as the Tibetan Plateau. Chapter 10 discusses the adjustment in Noah’s soil and vegetation parameters needed to reproduce the soil temperature states and surface fluxes measured at a site on the Tibetan Plateau. An analysis of the Noah simulations obtained using the standard parameterizations shows that, 1. The transfer of heat through the soil column underestimates the diurnal temperature cycle of the deeper layers; 2. The partitioning of solar radiation results in an overestimation of the sensible heat flux.

By using different soil thermal properties for the top- and subsoil, and including an additional thin top soil layer, the diurnal cycle of the temperature in deep soil layer is enlarged. This also improves the temperature simulation in the shallow soil layer. Further, a decrease in the vegetation parameters constraining transpiration was found to be necessary for reducing the sensible heat flux overestimation from 41 W m-2 to 20 W m-2. This illustrates once again that parameterizations utilized for global simulations can introduce large uncertainties locally, specifically over extreme environments, such as the Tibetan Plateau. Encouraging is, however, that model structures of land models, in this case Noah, are flexible enough to reproduce the measured land surface process with some minor adjustments.

Integration of satellite retrievals with land models

Some uncertainties within Noah related to its model structure and the utilized parameterizations were discussed in Chapters 9 and 10. Nowadays, a reduction in model uncertainties is also being pursued by integrating simulated with modeled land surface states, for which often data assimilation techniques are used. A fundamental assumption in the application of data assimilation techniques is that the modeled and retrieved soil moisture climatologies are unbiased. This is hardly ever the case. Moreover, the spatial resolution of LSM’s is at least several kilometers, while ASAR observes soil moisture at a 100 m resolution. A comparison of Noah simulations at a 10 km resolution against ASAR soil moisture retrievals from Chapter 8 yields, thus, deviations that are related to 1) the inherent bias due to differences in the climatology of the simulated and retrieved soil moisture, 2) the differences in the spatial representation, and 3) the uncertainties associated with the simulations as well as retrievals. An understanding of the contribution of these sources of difference is needed to fully appreciate the value of integrating satellite soil moisture with models operating at a lower spatial resolution.

Chapter 11 presents a method for quantification of these sources of differences. The decomposition of the sources of difference shows that during winters the spatial variability contributes for about 50% to the differences between the two products. This contribution increases to more than 70% for the monsoon affected period. In spite of the differences, however, the comparison of the two unbiased soil moisture products yields a RMSD of 0.051 m3 m-3 on a daily basis. This agreement between the simulated and ASAR retrieved soil moisture is no less than that for soil moisture products from coarse resolution microwave sensors.

As such, the integration of SAR based soil moisture products with large-scale simulations deserves more attention than it has obtained so far. One should, however, be aware that soil moisture differences between the large-scale simulation and high-resolution retrievals are largely caused by different spatial resolutions. Perhaps, the true value of using high resolution soil moisture products for large-scale modeling lies in improving the spatial soil moisture representation at the sub-grid level. Unfortunately, the temporal availability of SAR observations still hinders the development of operational high resolution soil moisture products.

PhD Defence Ceremony
PhD Thesis


previous page
more news

  1. Home »
  2. Organisation »