Soil-moisture nudging experiments with a single-column version of the ECMWF model

Q. J. R. Meteorol. SOC. (1999), 125, pp. 1879-1902

Soil-moisture nudging experiments with a single-column version of the ECMWF model

BY YUCHUN H U ~ , XLAOGANG G A O ~ , w. JAMES SHUTTLEWORTH]*, HOSHIN GUPTA~, JEAN-FRANCOIS MAHFOUF’ and PEDRO VITERB02 ’ University of Arizona, USA European Centre for Medium-Range Weather Fimcasts, UK

(Received 15 April 1998; revised 28 October 1998)

SUMMARY The soil-moisture nudging technique suggests using model forecast errors in near-surface air temperature

and relative humidity to re-initialize (update) soil moisture in atmospheric models. This study investigates the application of soil-moisture nudging using a single-column version of the European Centre for Medium-Range Weather Forecasts (ECMWF) model. The model was applied at 16 sites selected to sample a range of climates and land covers across the globe, with atmospheric forcing taken from the ECMWF operational analysis for 15 June 1994 and 15 December 1994. When observation errors are set to zero, the Optimal Interpolation technique for deriving estimates of nudging coefficients shows computational instability because strong correlation between near-surface temperature and near-surface relative humidity forecast errors makes the coefficient matrix of the linear equations ill-posed. Therefore, Principal Component Analysis (PCA) is used as a pre-processor to identify the independent and dependent components of temperature and relative humidity errors. With PCA, the soil- moisture nudging coefficients appropriate to the dominant principal component become stable and effective for correcting soil moisture within days. Such coefficients, when derived for the northern hemisphere summer over the 16 sites, are sufficiently consistent to propose using their all-site, daily-average values in a globally applicable soil-moisture analysis scheme. Tests of this method at the First ISLSCP (International Satellite Land Surface Climatology Program) Field Experiment (FIFE) site confirm that soil-moisture nudging can provide good estimates of near-surface weather variables and surface fluxes in a numerical weather prediction (NWP) model which gives poor simulation of precipitation (thereby poor soil moisture). However, the PCA method is unable to give an accurate determination of the location of the soil moisture within soil layers that are accessible to the atmosphere. Further, these tests show that, if the NWP model has good simulation of precipitation but poor simulation of surface radiation, soil-moisture nudging could wrongly vary the soil moisture so as to provide good estimates of the surface sensible-heat flux, but poor estimates of surface latent-heat flux. An approach to distinguish error sources from surface radiation and soil moisture is necessary for further improvement.

KEY WORDS: Data assimilation Land-surface processes Medium-range forecasting

1. INTRODUCTION

Atmospheric forecasts have been shown to be sensitive to the spatial distribution and temporal persistence of soil moisture (Rowntree and Bolton 1983; Shukla 1984; Oglesby 1991; Mintz and Walker 1993; Garratt 1993; Miliy and Dunne 1994; Gao et al. 1996). Therefore, weather and climate prediction is likely to be improved, not only by better representation of land-surface/atmosphere interactions in atmospheric models, but also by better initialization of soil moisture in the sub-models used to represent those interactions in numerical weather prediction (NWP) and general-circulation models (GCMs). Better initialization is arguably of most importance for short- and medium- range forecasts, i.e. for time-scales from days to seasons.

When using GCMs for long-term studies of climate, it is common to employ extended model ‘spin-up’ periods (typically of the order of several years) to minimize the simulation uncertainty associated with the poorly specified initial soil moisture. This strategy is clearly not suitable for short-term NWP models. Soil-moisture initialization is, however, known to have a significant impact on weather forecasts. Beljaars et al. (1996), for instance, found that perturbations in the initial soil-moisture settings of the

* Corresponding author: Department of Hydrology and Water Resources, University of Arizona, Tucson, AZ 85721, USA.

1879

1880 Y. HU et al.

European Centre for Medium-Range Weather Forecasts (ECMWF) land-surface scheme (Viterbo and Beljaars 1995) resulted in marked differences in the modelled precipitation 48-72 hours later.

Various strategies for improving the description of soil moisture in NWP models are currently under investigation. It is arguably possible to derive some benefit by improving the representation of surface-atmosphere exchange processes to give improved simulation of the seasonal development of soil moisture. Thus, replacing traditional simple ‘bucket’ representations of soil-moisture dynamics (e.g. Manabe 1969) with advanced Soil-Vegetation-Atmosphere Transfer (SVAT) schemes (e.g. Biosphere-Atmosphere Transfer Scheme; Dickinson et al. 1986) that simulate land-surface energy and water balance more comprehensively is likely to aid the simulation of soil-moisture evolution and therefore improve forecasts of weather parameters (air temperature, cloudiness, precipitation, etc.). However, to go further requires methods which re-initialize (update) the soil-moisture status to keep it close to reality.

It has been proposed that use of the model’s own SVAT scheme in a distributed, off-line mode with observed atmospheric forcing (radiation, precipitation) should result in improved simulations of the soil-moisture field, and that this synthetic soil-moisture field could be used for periodic updating of the predictive model. The National Center for Environmental Prediction (NCEP), for instance, plans to use observed precipitation and meteorological data fields in this way to update the Eta* model (Mitchell (1997), personal communication). Remotely sensed data, such as land-surface temperature (Jin et al. 1997) or microwave estimates of soil moisture (Houser et al. 1998), can, in principle, be used to provide estimates of soil moisture. However, there are still substantial technical issues to be resolved before the use of remotely sensed information in this role can become routine and widely applied. An alternative technique involves using the discrepancy between model-predicted and analysed observations of near- surface air temperature and relative humidity to adjust the values of the modelled soil moisture (Mahfouf 1991). This last method is the focus of interest in the present paper.

Mahfouf (1991) described the concept of soil-moisture nudging based on the Opti- mal Interpolation (01) technique and reported successful assimilation experiments with this technique for three, 48-hour periods with a one-dimensional version of the French Weather Service mesoscale model using data from the Hydrological Atmospheric Pi- lot Experiment (HAPEX-MOBILHY: Andre et al. 1988). Bouttier et al. (1993a, b) expressed nudging coefficients as functions of the time of day and of land-surface characteristics-specifically vegetation cover and aerodynamic roughness length. They applied these nudging coefficients to a three-dimensional mesoscale model for two days over a 400 x 400 km2 region in the HAPEX-MOBILHY study area. The results showed that, when a nudging technique was applied, soil moisture which was initially perturbed was restored towards values from a reference simulation within about 48 hours.

Viterbo (1996) introduced a simplified version of this technique into the ECMWF model. He set the soil-moisture adjustment equal to the bias in the low-level specific humidity modified by factors to take into account the fractional vegetation cover, the length of the analysis cycle (6 h), and a relaxation constant, D. He found that, with this simple nudging correction, a dry-warm drift of the ECMWF model coming from an underestimation of cloud cover was reduced. More importantly, the global Bowen ratio was reduced from an original value of 0.6-0.9 to 0.4-0.6, while the strong, summer influences of nudging on the atmosphere were found to extend to the entire troposphere. The effect of the nudging was found to be sensitive to the value assigned to

* ‘Eta’ is the grid-point primary regional model at the National Meteorological Center.

SOIL-MOISTURE NUDGING EXPERIMENTS 1881

the parameter D. Improved nudging coefficients along the lines of Mahfouf (1991) and Bouttier et al. (1993) could account for their geographical and time dependencies.

This study is a direct consequence of the need to extend the soil-moisture nudging method to global-scale application. Here, an investigation of the soil-moisture nudging concept is carried out. The basic hypothesis we test is that there is a relationship between errors in the forecast near-surface temperature and relative humidity and the error in the forecast soil moisture, and that this relationship can be successfully used to re-initialize soil-moisture fields in atmospheric models. We assume that the global-average strength of the relationship can be identified adequately in the face of other dependencies (e.g. seasonal influences, site-dependent factors, etc.) by carrying out single-column model simulations at many sites, providing that these sites are selected to sample a range of climates around the globe in winter and summer conditions.

In the next section, we apply the 01 method for calculating soil-moisture nudging coefficients (Mahfouf 1991) to 16 sites selected from different climatic regimes across the globe. After analysing the features of the ensuing nudging coefficients, we propose to use a Principal Component Analysis (PCA) pre-processor to improve the stability of the calculated nudging coefficients and then compare these two approaches. We then demonstrate the effect of applying nudging at the 16 selected sites using a single set of all-site-average, all-day-average nudging coefficients. In section 3, this set of nudging coefficients is applied in a 3-month test of nudging at the First ISLSCP (International Satellite Land Surface Climatology Program) Field Experiment (FIFE) site. The summary and conclusions are provided in section 4.

2 . MODELS AND METHODS

( a ) Model description The model used to investigate the relationship between the forecast errors in soil

moisture and the associated errors in screen-level temperature and relative humidity is a single-column model of the ECMWF weather prediction model (hereafter called the SCM). Only the vertical diffusion represented in the original ECMWF model is retained in the atmospheric portion of the SCM. In the land-surface portion, the SVAT scheme described by Viterbo and Beljaars (1995) is used. This includes representation of heat and water transfers over four soil layers (7 cm, 21 cm, 72 cm, and 189 cm). Different types of land surface are classified using roughness length and vegetation fractional coverage. To improve the calculation of exchanges at the atmosphere-land interface, a vegetation canopy without heat capacity is overlaid on bare ground. The atmosphere and the underlying land system interact within the SCM by the exchanges of latent- and sensible-heat fluxes through a planetary boundary layer sub-model that provides the interface between the lowest modelled layer in the atmosphere and the skin layer. The SCM is forced by surface solar radiation and precipitation and by the prescribed time- varying horizontal convergence of water vapour, heat and wind from ECMWF analyses.

In the SCM, both the temperature and relative humidity at screen height (2 m above surface) are diagnostic variables which are obtained by interpolating between the lowest layer of atmosphere (about 30 m above the surface) and the skin layer from the Monin- Obukhov similarity theory. The temperature and humidity at screen height are sensitive to variations in the skin layer which are in turn affected by the water and energy exchanges modelled by the land-surface scheme. These latter exchanges are directly related to the soil-water contents in the upper layers, and the nudging effect stems from these dynamic relationships. Thus, the method described hereafter uses detectable

1882 Y. HU et al.

forecast errors in screen temperature and relative humidity as indicators of modelled errors in soil-water content.

(b) Principle of soil-moisture nudging

(i) General description. The soil-moisture nudging principle is based on the 01 technique (e.g. Lorenc 1986). The problem addressed by the 01 method is that of finding the optimal analysed value of a variable from observed and predicted values of that variable. The 01 method produces a solution (analysis) that equals the predicted value plus the weighted difference between observed and predicted values, with weights determined by minimizing the root-mean-square error between analysed and true values.

Mahfouf (1991) proposed adopting the 01 method for simplifying the complex relationship between the errors of soil moisture and errors of temperature and relative humidity via the linear expression:

where T = the air temperature at screen height, R H = the relative humidity at screen height, W = the soil-water content, and cq, = 'nudging coefficients' (to be determined).

The subscript 's' refers to the surface soil layer (in the present study, 0-7 cm), while subscript 'd' refers to the deep soil layer (in the present study, 7-100 cm) represented in the SCM. The superscript '0' denotes observed values of temperature and relative humidity, while superscript 'p' denotes the predicted (first guess) values of temperature and relative humidity. Similarly, the superscripts 'p' and 'a' denote predicted and analysed values of soil moisture in the two soil layers, respectively.

The nudging coefficients a1,2 (similarly for /?1,2) can be solved from the following system:

where OX = the standard deviation of variable X and p x , y = the correlation coefficients of variables X and Y .

Equation (2) (similar to Eq. (23) in Mahfouf (1991)) indicates that obtaining values of the coefficients requires that the statistics of observation and forecast errors are known. The observation errors for air temperature and relative humidity can be estimated with reasonable accuracy from instrumental errors and representativeness errors. In the Mahfouf (1991) study, the standard deviation of observation error for T was set to 2 K, while the standard deviation of observation error for R H was set to 15%. Knowledge of forecast errors is more difficult to obtain. In theory, the forecast errors should include both model (structure and parameters) errors and initial soil-moisture errors. Under the condition of relatively small model errors, the initial soil-moisture errors can be derived from empirical models. The technique adopted in this work is based on a Monte-Carlo approach described by Mahfouf (1991).

First, a model run, the so-called 'control run', is made with an initial soil moisture in the SCM set to the middle of the range allowed in the land-surface scheme that is


under investigation. The output from this model run is designated to be a substitute for the true values of the atmospheric screen-level parameters. An additional set of model runs is then made with perturbed initial soil-moisture contents. Analysis of the difference between the ‘control’ and ‘perturbed’ time series of modelled near-surface weather variables and soil moisture provides information on forecast errors in T and R H produced by soil-moisture errors. Thus, the derivation of Eq. (2) includes several assumptions regarding both forecast errors and observation errors which make it an approximation of Eq. (1) .

When using ‘simulated observations’ (so-called ‘twin experiments’), it is sometimes convenient to study the influences of the forecast errors in T and R H on the calculation of nudging coefficients by using the simplified version of Eq. (2) given in the limit of zero observation errors. The simplified form of Eq. (2) is:

with the solution:

(ii) Methodology. In this study, the ECMWF SCM was run at 16 sites which were selected to represent a range of climatic regimes and land-surface types across the globe that includes tropical forest, temperate prairie, tropical and middle-latitude desert, boreal forest, etc. The geographical location of the selected sites is shown in Fig. 1, while Table 1 provides site-specific information on their latitude and longitude and the fractional vegetation cover and aerodynamic roughness for momentum, which is prescribed for each site within the ECMWF land-surface scheme. Two sets of runs were made, one in the northern hemisphere summer and one in the northern hemisphere winter, with the site-specific forcing data for the SCM taken from the ECMWF observational analysis on 15 June and 15 December 1994, respectively.

At each site and each time of year, the control run and 100 perturbed runs were made using the same forcing and with the same daily forcing ‘recycled’ three times to produce a set of 3-day model calculations of atmospheric and land variables for statistical analysis. The values of the (generic) variables X and Y calculated in the control and perturbed runs on the third day were used to calculate the statistics of OX and px,y at a specified time of a day (Note: N (number of runs) = 100 in this study). In the perturbed runs, the initial soil moisture was randomly chosen from within the range (Wwilt + Wcap)/4 to 3( Wwilt + Wcap)/4, where Wwilt and Wcap are the volumetric water content of the soil at wilting point and field capacity respectively, prescribed in the ECMWF land-surface scheme. The runs were all made with precipitation processes suppressed in the SCM model and with the land-surface scheme driven by surface solar radiation taken from ECMWF short-range forecasts. In this way, possible errors due

1884 Y. HU et al.

Figure 1. Location of the 16 sites used in this study which are selected to sample a range of climates and land cover across the globe.

TABLE 1. GEOGRAPHICALLOCATION OF THE 16 SITES USED IN THIS STUDY AND THE SURFACE PARAMETERS RELEVANT AT THESE SITES AS SPECIFIED

WITHIN THE ECMWF LAND-SURFACE SCHEME

Fractional Aerodynamic Site Latitude Longitude vegetation cover roughness length name (ON) (“E) (dimensionless) (m)

ARME Australia BOREAS Cabauw CART EFEDA FIFE HEIFE MAGS MOBILHY MONSOON NOPEX Sahara Sahel Siberia Tibet

-2.5 -23.9

55.3 52.5 37.3 39.6 39.6 41.3 65.4 44.1 15.4 60.4 22.2 16.6 65.4 32.3

299.8 133.9 252.2

5.0 262.0 356.7 262.0 99.6

235.0

74.8 17.0 1.7 2.3

100.0 95.0

0.75

0.94 0.44 0.87 0.90 0.82 0.80 0.84 0.43 0.88 0.84 0.83 0.84 0.43 0.44 0.85 0.84

2.50 0.10 1.50 0.24 0.24 0.72 0.27 0.33 0.46 0.68 0.5 1 1.60 0.01 0.11 0.89 0.56

Some of the sites are designated by acronyms corresponding to field experiments, as follows. ARME refers to the Amazon Regional Micrometeorological Experi- ment; BOREAS refers to the Boreal Ecosystem-Atmosphere Study; CART refers to the Cloud And Radiation Testbed of the Atmospheric Radiation Measurement Experiment; EFEDA refers to the ECHIVAL (European project on Climatic and Hydrological Interactions between Vegetation, Atmosphere and Land Surface) Field Experiment in a Desertification-threatened Area; FIFE refers to the First ISLSCP (International Satellite Land Surface Climatology Program) Field Exper- iment; HEIFE refers to the Hei-He River Field Experiment; MAGS refers to the Mackenzie GEWEX (Global Energy Water-cycle Experiment) Study; MOBILHY refers to the HAPEX-MOBILHY (Hydrologic-Atmospheric Pilot Experiment- Modelisation du Bilan Hydrique); MONSOON refers to the MONSOON’90 experiment; and NOPEX refers to the Northern hemisphere Pilot Experiment.


to poor simulation of rainfall and radiation when the ECMWF model is used in this one-dimensional format are not taken into account in the forecast errors, but the land- atmosphere feedback on near-surface meteorological variables given by the modelled surface heat fluxes is maintained.

(iii) Preliminary results. The nudging coefficients were calculated from both Eq. (2) (with observation errors) and Eq. (4) (without observation errors) for each site, for each hour, and in both the northern hemisphere summer and winter seasons. Figure 2 shows the results for (2'1 and (2'2, the nudging coefficients calculated for the surface soil layer, for two (of the 16) sites. The nudging coefficients for the deep soil layer, #I1 and p2 (not shown), exhibit behaviour similar to that of a1 and a2, but they are larger. In general, the calculated values of the coefficients exhibited marked variation throughout the day, and they changed in a dramatic way between sites and between seasons at individual sites. Not only do the nudging coefficients change in magnitude, but they also change in sign.

A careful inspection of Fig. 2 illustrates the source of the problem. If, for instance, the soil is predicted to be drier than the observation, in most cases the screen temperature is expected to be biased positively and the relative humidity to be biased negatively. Thus, a1 and (2'2 should have opposite signs, and Eq. (I) will then include two terms with the same sign, one corresponding to temperature and the second to relative humidity, both tending to nudge the soil moisture in the same direction. The sum of these two terms provides the correction to the soil-water content. However, Fig. 2 shows that the derived values of (2'1 and (2'2 do not always have opposite signs. Moreover, they exhibit apparently haphazard, simultaneous shifts (both more positive or both more negative) from one hour to the next. These shifts compensate for each other in Eq. (I). Thus, the soil-moisture nudging is not altered by the presence of this simultaneous, haphazard contamination of the two nudging variables because this is related to the difference between the two nudging terms, but the absolute magnitude of the two nudging coefficients is poorly defined because of them. The contamination is presumably determined by the details of the forcing conditions for which the nudging coefficients are derived. However, the contamination appears to be arbitrary and it can be large, and its presence precludes attempts to provide a simple parametrization of their value (e.g. in terms of surface cover or aerodynamic roughness) using globally applicable formulae.

The above results indicate that the fundamental cause of the lack of stability in the calculated values of the nudging parameters is the high correlation between the forecast errors of near-surface temperature and relative humidity. The calculated correlation coefficients between T and R H from model runs at the 16 study sites calculated for four times during the day and for both northern hemisphere summer and winter are given in Table 2. As expected, in the northern hemisphere summer, T and R H exhibit high negative correlation at all 16 sites at all times of day. In the winter, the situation becomes complicated. Low-latitude sites, those that lie between 30"N and 30"S, i.e. the ARME, Australia, MONSOON, Sahara, and Sahel sites, still show a high negative correlation between T and R H . High-latitude sites, those that lie above 60"N, e.g. MAGS and Siberia, exhibit a reduced correlation. (Note: At some sites, e.g. NOPEX, the atmosphere saturates in the model runs at this time of year.) At mid latitudes, some sites (EFEDA and MOBILHY) show negative correlation, while others (Cabauw and HEIFE) show a weak positive correlation. Still others (FIFE and Tibet) show positive correlation at some times of day and negative correlation at other times. On the basis of these results, it seems evident that the presence of a strong surface-energy forcing has a major influence on the degree of correlation between T and R H forecast errors.

1886

-0.02

-0.04

Y. HU et al.

.

.

x 10'' I 0.04 I 1

h r

k

5 E v

1 6 12 18 24

x 10~' 31 1

I ' I ' I' II

-0.06 12 18 24

0.01 , 1

2.5 1 n 2

E 1.5 v 0 ( u 1

0.5

0

- I

-0.005

tj

-0.5 I 1 6 12 18 24

x 10"

-2 ' I 1 6 12 18 24

x 1 0 - 4

-0.01 1 1 6 12 16 24

x 10.' 2

1.5

1

0.5

0

-0.5

1 1 6 12 18 24

x lo-* 2.5 r "I I

2

1.5

1

0.5

0

I 1 6 12 18 24

Local Time (hours) -0.5 ' 1

1 6 12 i a 24

Local Time (hours)

Figure 2. Values of nudging coefficients a1 and a2 at two study sites derived for each hour of the day using the Optimal Interpolation method for the northern hemisphere summer ((a), (c), (e), and (g)) and for the northern hemisphere winter ((b), (d), (t), and (h)). The sites shown are the (mid-latitude) FIFE site (- x -) and the (low latitude) Sahara site (- o -). Figures (a), (b), (c). and (d) are calculated for the limiting case in which observational errors are zero, while Figs. (e), (0, (g), and (h) are calculated assuming observational errors of 2 K for temperature

and 15% for relative humidity. See Table 1 for details of the sites.

SOIL-MOLSTURE NUDGING EXPERIMENTS 1887

TABLE 2. VALUE OF THE CORRELATION COEFFICIENTS BETWEEN AIR TEMPERATURE AND RELATIVE HUMIDITY AT 2 M FOR THE STUDY SITES DURING NORTHERN HEMISPHERE SUMMER AND NORTHERN

HEMISPHERE WINTER

Forecast time in Forecast time in northern hemisphere summer northern hemisphere winter

Site name 06:OO 12:OO 18:OO 24:OO 06:OO 12:oo 18:OO 24:OO

ARME Australia BOREAS Cabauw CART EFEDA FIFE HEIFE MAGS MOBILHY MONSOON NOPEX Sahara Sahel Siberia Tibet

-0.995 -0.992 -0.997 -0.987 -0.998 -0.987 -0.996 -0.993 -0.993 -0.994 -0.999 -0.998 -0.993 -0.996 -0.930 -0.999

-0.992 -0.989 -0.980 -0.998 -0.974 -0.999 -0.996 -0.993 -0.994 -0.992 -0.993 -0.991 -0.994 -0.989 -0.995 -0.993 -0.991 -0.987 -0.994 -0.991 -0.999 -0.994 -0.996 0.992 -0.997 -0.793 -0.985 -0.961 -0.998 -0.997 -0.997 -0.996

-0.992 -0.992 -0.999 -0.992 -0.993 -0.991 -0.989 -0.993 -0.984 -0.994 -0.999 -0.997 -0.994 -0.987 -0.996 -0.998

-0.990 -0.992 -0.813

-0.949 -0.990 -0.965

0.830

0.982 -0.863 -0.978 -0.994

-0.919 -0.918 -0.599

0.018

-0.992 -0.995

0.822 0.244

-0.961 -0.956

0.988

-0.976 -0.995 -0.847 -0.940 -0.992 -0.61 1

0.806

-0.974 -0.994 -0.937

-0.970 -0.993 -0.278

0.962

0.906

-0.998 -0.995

-0.974 -0.992 -0.699 -0.984

-0.985 -0.996 -0.914

0.933 -0.997 -0.994

0.102 0.905

-0.998 -0.996

-0.964 -0.992 -0.904 -0.872

Note: Missing values occur when atmospheric saturation occurs at a site and correlation coefficients are consequently not defined.

In the absence of observation errors, Eq. (4) clearly shows that the solutions a 1 and a 2 can exhibit dramatic variations in value whenever:

In many of the cases given in Table 2, the correlation coefficient pT, RH x - 1. Table 2 also implies pW,T x - p W , R H . The dramatic changes in sign and magnitude of the calculated original coefficients are therefore to be expected. They are merely a consequence of the frequently high correlation between near-surface temperature and humidity forecast errors induced by soil-moisture errors when surface energy is a major control on local, near-surface weather variables. In Fig. 2, the comparison between a1 and a 2 with and without observation errors demonstrates that including observation errors tends to mitigate, but not totally remove, the contamination. This is to be expected from Eqs. (2) and (3).

The marked negative correlation between near-surface temperature and relative humidity forecast errors in conditions of strong surface-radiation forcing implies that the nudging method could be redefined by accounting for such correlation explicitly. The alternative use of a PCA as a pre-processor to overcome this problem is now explored.

(c ) The Principal Component Analysis method The application of PCA to Eq. (1) involves the definition and use of two new

variables, Y1 and Y2. These two variables are the linear combinations of the original variables T and R H which capture their correlated and uncorrelated components and

1888 Y. HU et al.

which are created by a process of orthogonal linear transformation, thus:

with the restrictions that py, y2 = 0, and that:

where 19 is the angle of rotation of the coordinate axes required for an orthogonal linear transformation.

Notice that, in Eq. (5 ) , temperature and relative humidity are normalized relative to their respective mean values and standard deviations to give the dimensionless variables T* and RH*. This is to ensure that the rotation matrix, u i j , is not affected by the units of temperature and relative humidity. Hereafter, the analysis is made using those normalized values.

Substituting Y1 and Y 2 into an equation equivalent to Eq. (1) gives:

Without observation errors, this has a solution which is similar to Eq. (4), but simplified by py,,y2 = 0, i.e.:

where the forecast errors are:

With observation errors, the solutions are:

where the observation errors are:

(Note: W, and w d have the units metres of liquid water, while a 1 , 2 and /31,2 have units metres-' .)

The use of the PCA method provides several important advantages over the 01 nudging using temperature and relative humidity directly, as follows:


(1) The first principal component (Yl ) of T and R H explains most of the variance in these two variables (99.3%-99.9% of the variance in summer and 84.9%-99.9% in winter). Hence, the second principal component (Y2) is negligible in most cases, which reduces the total number of nudging coefficients from four (a1 , a 2 , PI, and 82) to three (a1 , 61, and 0). (Note: If both principal components are retained, there is obviously an inverse transfer between the PCA nudging coefficients and the original 01 coefficients.)

( 2 ) In conditions and at locations where surface heating of the atmosphere is high, the PCA nudging coefficients vary less than the original coefficients and have constant sign during their diurnal cycle.

(3) Using the PCA approach, the nudging coefficients with observation errors can be calculated in two steps: first, the coefficients without observation errors are calculated using Eq. (7); then they are multiplied by factors determined for the prescribed values of observation errors (Ey. (8)).

Accordingly, the approximate formula using only the first principal component is:

(Note: In the derivation of Eq. (Z), the mean of the predicted values is assumed to be equal to the mean of the observed values; then both the means of T and R H are eliminated in Eq. (9).)

Equation (9) captures the physical meaning of the PCA nudging method. Only the dominant (>99% variance in summer) parts of T and R H are used to relate to the soil- moisture contents; the ratio between a1 and represents the relative distribution of nudging water between the surface and deep soil layer, and v11 and 1112 represent the weighted contribution of T and R H to the soil-moisture nudging (the opposite signs of u11 and u12 correspond to the opposite correlation between soil-moisture errors and the errors in T and R H ) .

Figure 3 illustrates diurnal variation of the nudging coefficients, a; and u11, used in the PCA nudging method (pi is similar to a; in form but with a larger absolute magnitude). Figures 3(a) and (b) show that, in the northern hemisphere summer, the PCA nudging coefficients have the same sign and similar magnitude at all the sites, regardless of locations and land-cover characteristics. This suggests (which will be explored later) that a single set of all-site, daily-average values for the PCA nudging coefficients might suffice when surface energy fluxes are high and soil-moisture nudging is, consequently, legitimate.

Figures 3(c) and (d) show that, in the northern hemisphere winter, the behaviour of the PCA nudging coefficients is much less consistent. In this case, the sign of the PCA nudging coefficient a; varies with latitude. For low-latitude sites (those lying between 30"N and 30°S, e.g. the ARME and Sahara sites) a{ retains the same sign as in the northern hemisphere summer. For high-latitude sites a; changes sign relative to its northern hemisphere summer value. However, in the case of mid-latitude sites, some (e.g. the Cabauw site) keep northern hemisphere summer features like the low-latitude sites and some (e.g. the BOREAS site) behave like high-latitude sites. Figure 3(d) shows that, in the case of the nudging coefficient u11, there are frequent and worrisome changes of sign at some sites in the northern hemisphere winter.

1890

0.04

0.03

0.02

0.01

” - t ) O

-0.01

-0.02

-0.03

-0.04

1

0.8

0.6 h

T

1 0.4 E v

0.2 - r > o

-0.2

-0.4

-0.6

-0.8

-1

Y. HU ef al.

0 10 15 20 25 5

- ARME -. BOREAS * CABAUW x FIFE -- SIBERIA -0 SAHARA

(b)

5 10 15 20 25 LOCAL TIME (hours)

0.04

0.03

0.02

0.01

0

-0.01

-0.02

-0.03

-0.04

1 n 5 10 15 20 25

LOCAL TIME (hwrr)

Figure 3. Values of nudging coefficients at six study sites (see Table 1 ) derived for each hour of the day using the Principal Component Analysis method described in the text for the northern hemisphere ((a) and (b)) summer

and ((c) and (d)) winter.

Table 3 summarizes the above results by providing all-site, daily-average values of the PCA nudging coefficients in the northern hemisphere summer and winter, with sites appropriately grouped by latitude in the case of winter. The values given in Table 3 for the northern hemisphere winter at mid latitudes and high latitudes are certainly of limited use. During winter, the correlation between soil-moisture errors and near-surface weather variables is low because of the low radiative forcing and because soils are close to field capacity; hence, surface evaporation is influenced little by soil moisture.

( d ) Comparison of application of the 01 and PCA methods A comparison was made between applications of nudging using the 01 and PCA

nudging coefficients at sites for which the coefficients were derived assuming zero


TABLE 3. DAILY-AVERAGE PRINCIPAL COMPONENT ANALYSIS NUDGING COEFFI- CIENTS (SEE TEXT) AVERAGED FOR ALL SITES IN NORTHERN HEMISPHERE SUMMER AND FOR GROUPS O F SITES FOR NORTHERN HEMISPHERE WINTER WITHOUT OBSERVATION

ERRORS

Season Site a; (m-') S ; (m-') v t i UT (K) ~ R H (%)

Summer All sites 0.0026 0.0558 -0.6186 2.657 15.44 Winter Low-latitude 0.0044 0.0663 -0.5382 3.575 19.62 Winter Mid-latitude 0.0038 0.0657 -0.4374 0.686 11.60

0.0579 0.7429 0.686 11.60 Winter Mid-latitude 0.0035

Winter High-latitude -0.0030 -0.0506 -0.9645 2.591 0.091

(1)

(11)

Note: Low-latitude sites (3OoS-30"N) include ARME, Australia, Sahara, Sahel, and MON- SOON. Mid-latitude sites (30°S-600S and 30"N-60°N) in group (I) include CART, EFEDA, MOBILHY, and Tibet, and in group (11) include Cabauw, HEIFE. High-latitude sites (60"s- 90"s and 6O0N-90"N) include Siberia.

observation errors. Comparative soil-moisture nudging experiments were made at all 16 study sites in summer and winter using values of 01 and PCA nudging coefficients which included their original, site-specific, diurnal variations. In these tests, the value of soil moisture was initially different from that used in the control run (as it was for all perturbed runs). However, every six hours (at OO:OO, 06:00,12:00, and 18:OO local time), the modelled values of near-surface temperature and relative humidity were compared with the corresponding values given by the control run (recall that values calculated by the control run are substitutes for observations). The errors in the model-calculated values of T and R H given with an initially perturbed soil moisture at the time of each comparison were substituted into Eq. (1) (the 01 method) or Eq. (8) (the PCA method) to calculate the soil-moisture errors. The soil-moisture state in the model was then updated using the corresponding estimated error in soil-moisture state given by these equations. Meanwhile, the atmospheric state was also changed to that of the control run at the equivalent time.

Example results of the above-described nudging experiments in the summer at three sites are displayed in Fig. 4 for the ARME, FIFE, and Sahara sites (along with the results of similar nudging experiments with all-site average values of nudging coefficients described below). In all cases shown in Fig. 4, the nudging process works more effectively on the first day than on subsequent days. In general, there is little obvious difference in the effectiveness of the original 01 and PCA methods when the surface energy exchange is large at all sites in the northern hemisphere summer and at the low-latitude sites (ARME and Sahara) in both the summer and winter.

The need for a nudging technique to be applicable in a global NWP model is a critically important criterion for evaluating the relative merit of alternative nudging techniques. For the nudging coefficients derived using the PCA technique, the northern hemisphere summer nudging coefficients at all 16 sites have the same sign and exhibit small changes in magnitude from site to site and from time to time. As mentioned earlier, this suggests the possibility that a single set of all-site, daily-average, northern hemisphere summer nudging coefficients might be adequate for use in soil-moisture nudging. To test this, we applied the appropriately averaged coefficients listed in Table 3 in nudging tests at all 16 study sites. Satisfactory soil-moisture convergence was obtained in all 16 cases. Figure 4 shows the comparison for three sample sites. There is little obvious difference between the soil-moisture nudging given by using all-site,

1892

0.35

Y. HU el a1

a .

0.15

0.1

0.35 I

’

’

d l

0.1 0’15 I 0.05

0 12 24 36 48 60 72

Time (hour) 0 12 24 36 48 60 72

Time (hour)

C

f

0 12 24 36 48 60 72

Time (hour)

Figure 4. The effect of nudging at 6-hour intervals for 3 days using site-specific, time-specific values of nudging coefficients derived using the Optimal Interpolation (01) method and using the site-specific and all-site- average Principal Component Analysis (PCA) methods described in the text for the northern hemisphere summer. Figures (a) and (d) are for the ARME site, (b) and (e) are for the FIFE site, and (c) and (f) are for the Sahara site (see Table 1). Figures (a), (b), and (c) describe changes in the soil moisture of the surface layer, while (d), (e), and (f) describe changes in the soil moisture of the deep soil layer. The results shown are for the control run (- . -), the perturbed run without nudging (. . . ), the perturbed run with nudging using 01 coefficients (dark solid lines), the perturbed run with nudging using site-specific PCA coefficients (grey solid lines), and the perturbed run with

nudging using all-site-average PCA coefficients (dashed lines).

daily-average values of the PCA nudging coefficients and that given by the site-specific 01 and PCA coefficients.

An effectiveness parameter, S j , was defined to provide a quantitative measure of the relative effectiveness of the nudging process using the three nudging processes on the j th day ( j = 1 , 2 or 3), as follows:

WP - Wcontrol 2 JCi=1,2,3,4(6h)( I 1 ) 6 . - 7

J - control ~WPnitial- Winitial I

The values of Sj averaged over 16 sites in the northern hemisphere summer are given in Table 4. On average, the PCA methods are slightly less effective than the 01 method when site- and time-specific values of nudging coefficients are used in the conditions in which they were derived. This is because PCA does not use all of the variance in the T and R H errors. There is then a difference between using all-site average PCA nudging coefficients and site-specific, all-day average PCA nudging coefficients also.

(e) Comparison of PCA method and nudging using a single parameter (T or RH) The fact that temperature and relative humidity forecast errors produced by soil-

moisture errors are frequently highly correlated raises the possibility that soil-moisture nudging may be as efficient using just one of these variables rather than the first principal


TABLE 4. COMPARISONS O F THE EFFICIENCY O F THE OP- TIMAL INTERPOLATION (01) AND PRINCIPAL COMPONENT ANALYSIS ( P C A ) NUDGING METHODS AS MEASURED B Y THE PARAMETER 8, (SEE TEXT FOR DEFINITION) AVER-

SPHERE SUMMER FOR THE THREE SUCCESSIVE DAYS OVER AGED OVER ALL 16 STUDY SITES FOR THE NORTHERN HEMI-

WHICH SOIL-MOISTURE NUDGING IS APPLIED

Northern hemisphere summer 81 (%I 82 83 ("/.I

0 1 method 64.8 f 4.2 22.3 f 3.4 14.3 f 2.3 PCA site-specific 66.8 =t 4.4 25.9 * 3.5 18.0 i 2.2 PCA all-site 65.0 f 2.8 20.7 f 4.2 15.8 f 1.9

E 0.35 a c cn .- r"

0.3. v) a s 5 0.25. v)

0.2.

0 6 1218243036424854606672 Time (hour)

, . , I , , . , , , , , , , , )

0 6 1218243036424854606672 Time (hour)

Figure 5. The effect on surface soil moisture (m) of nudging at 6-hour intervals for 3 days using three types of site-specific, time-specific nudging coefficients at the FIFE site (see Table 1) for (a) northern hemisphere summer (surface layer), and (b) northern hemisphere summer (deep layer). The results shown are for the control run (---), the perturbed run without nudging (--), the perturbed run with nudging using PCA coefficients (- o -), the perturbed run with nudging using temperature only (- x -), and the perturbed run with nudging using relative

humidity only (- * -).

component of the two. To investigate this, we carried out a nudging test at the FIFE site using site-specific and time-specific nudging coefficients derived using a PCA analysis and compared this with nudging using coefficients derived for temperature and relative humidity separately. Such simplification is similar to the nudging process by Viterbo (1996), where only specific humidity is used. The results of this test are shown in Fig. 5. In summer, at the FIFE site, the results of nudging using these different variables are similar, thus confirming the value of the approach used by Viterbo (1996) in such summertime conditions.

3. TESTS OF PCA NUDGING AT THE FIFE SITE

A practical application of the PCA nudging method was investigated in a real-case situation at the FIFE site by exploiting the data available at that site for the period from 1 June to 30 August 1987. In these tests, the required atmospheric forcing was taken from the ECMWF re-analysis (Betts et al. 1998), but observed surface radiation and

1894 Y. HU et al.

precipitation were substituted for the model-calculated values. A preliminary test was made of the SCM’s ability (model accuracy) to describe the meteorology and hydrology at the FIFE site during this period when forced in this way. Figure 6 illustrates the comparison between soil-moisture content of the surface and rooting-zone soil layers and the model-calculated sensible- and latent-heat fluxes given by this control run and observed data. The results from the SCM modelling system used in this study confirm the results of Viterbo and Beljaars (1995) obtained in a stand-alone mode, and show similar, good agreement between the observed and modelled soil moisture during the entire period. However, they also show errors in the forecast surface heat fluxes, for instance, between Julian days 200 and 215, when the modelled latent heat exceeds observed values by about 50 W m-2, and there is a complementary underestimate in sensible-heat flux. These errors result largely from weaknesses in SCMs that cannot be corrected using nudging techniques.

Two experiments were carried out using the PCA nudging coefficients including observation errors. The observation error assumed for temperature was 2 K and that for relative humidity was 10%. These observation errors and the northern hemisphere summer, all-site, daily-average values of PCA nudging coefficients given in Table 3 were substituted into Eq. (8) to give the following set of the northern hemisphere summer, all- site, daily-average (hereafter called ‘global’) values of the PCA nudging coefficients:

ar(globa1) = 0.00359; /?;’(global) = 0.0557; v; ’~ (global) = -0.6186.

In these experiments, observed values of air temperature and relative humidity are used to nudge soil moisture.

(a) Experiment I: Erroneous rainfall The first experiment involved artificially changing the precipitation input used in

the model. A comparison was made between the simulation given in the control run and that given with rainfall from 1 June to 15 July arbitrarily set equal to zero and with the rainfall triple the observed value from 16 July to 3 1 August. Two perturbed runs were made, one with no application of soil-moisture nudging, and a second with soil-moisture nudging applied using the global values of PCA nudging coefficients given above.

The results of Experiment 1 are shown in Fig. 7. Figure 7(a) shows the time series of the observed daily total precipitation over a 92-day period from 1 June to 3 1 August (i.e. Julian days 151-243) and Fig. 7(b) shows the artificially changed rainfall pattern used in the perturbed runs. Figures 7(c) and (d) illustrate the modelled soil moisture in the surface and the deep layer, respectively, while Figs. 7(e) and (f) show the modelled sensible- and latent-heat fluxes, respectively.

In the following discussion, the period of Julian days 151-196 is referred to as Period I, while 197-243 is referred to as Period 11. The figures clearly indicate that the agreement with artificially perturbed precipitation patterns is greatly improved by the application of soil-moisture nudging in both periods and at the transition between them. The nudging becomes most active when there is observed rainfall (Period II), because this is when there is the most divergence between the control runs and those with modified rainfall. Clearly, continuous nudging has the potential to compensate for the error in the modelled soil moisture induced by wrongly predicted precipitation in an NWP model. However, it is important to note that, in Fig. 7(d), the deep- layer soil moisture during Period I is ‘over-nudged’, while that in the surface layer is ‘under-nudged’ . The near-surface temperature and relative humidity are sensitive to the available moisture in both soil layers. The use of different coefficients a1 avd


+ ( + Observation - Model) +

0.1 I I I I I I I I

150 160 170 180 190 200 210 220 230 240 250

g 0.35

0 v)

al al

Q 0.2 -

n 0.15 I I I I 1 I I I I wF 150 160 170 180 190 200 210 220 230 240 250

50

+ c Q O

I" al

v) C al

-50 .-

(I) -100 150 160 170 180 190 200 210 220 230 240 2

n N

0

3 -50

I€

3 -100

v c

I E -150 al (D c -J -200

150 160 170 180 190 200 210 220 230 240 250 Julian day

Figure 6. Output from the control run and observed data for the FIFE site (see Table 1) for the period from 1 June to 31 August 1987 (Julian days 151-243). The variables illustrated are (a) soil moisture in the surface layer (m),

(b) soil moisture in the deep layer (m), (c) surface sensible-heat flux, and (d) surface latent-heat flux.

for the surface and the deep soil layers should, in principle, enable a correct splitting of the atmospheric increments. For example, over bare soil the main source of moisture comes from the surface reservoir. However, the use of globally averaged coefficients, including both bare soil and vegetated locations, may prevent a correct partition. Moreover, for vegetated areas in the ECMWF land-surface scheme, water can be extracted from the surface reservoir through both plant transpiration and bare- soil evaporation, which can also make it difficult for the nudging method to distinguish bcpeen the two alternative influences. However, Figs. 7(e) and (f) show that nudging overall &',moisture in the perturbed model is very effective in correcting the modelled value of he s;psible- and latent-heat fluxes. Presumably, this is because of the strong, physically based &k between these fluxes and near-surface meteorology, and the cause

1896 Y. HU el al.

0 300 p 200- - C

p 0 8 a loo- 8 0

I I I I I I I I 1 -

(4 - Period - II Period - I

-

I nin - n na n I I 4 - 1 d. r I

0.4 e!

$ 0.3 si 0.2

t z 0.1

C

z - 300 $200-

: 100- " +

1 0

I I I I 1 1 I I I I

150 160 1 70 1 80 190 200 21 0 220 230 240

I 1 1 I 1 I I I I

(b) - without rain triple rainfall

-

I I I I " I I Il l n R I _

2 0.35 2 0.3 0.25

0 0.2

v1

-

0 0.15 I I I I I I I I

150 160 170 180 190 200 210 220 230 240

T o t r" -50 a -

f -loo

150 160 170 180 190 200 21 0 220 230 240

0 m- i -50 5-100 I $ -150 -200 I I I I I I I I I I

1 50 160 1 70 180 190 200 210 220 230 240 Julian day (1 987)

Figure 7. Outputs from Experiment 1 at the FIFE site (see Table 1) for the period from 1 June to 31a&gust 1987 (Julian days 15 1-243). The variables illustrated are (a) observed precipitation, (b) modified R2cipitation, (c) soil moisture in the surface layer (m), (d) soil moisture in the deep layer (m), (e) surface Sell$ble-heat flux, and (0 surface latent-heat flux. In each case, the full line is the model output for the control ru;, the dashed line

is for the perturbed run without nudging, and the line with circles is for the perturbed qi with nudging.


of the above-mentioned weakness in the definition of the source of moisture available to the atmosphere is not relevant for screen-level humidity and temperature (any lack of moisture in one soil layer is compensated for by additional moisture in the other soil layer).

In summary, on the basis of this experiment, soil-moisture nudging using global values of nudging coefficients derived using the PCA technique seems capable of providing a useful improvement in modelled surface fluxes and, consequently, near- surface weather variables when precipitation is unrealistically simulated. However, the nudged soil-moisture status is, of course, model dependent, and the distribution of soil- moisture increments between individual soil layers can be difficult for some land-surface schemes.

(b) Experiment 2: Erroneous surface radiation The second experiment is similar to the first except that, in this case, the precipitation

was as observed, but surface incident solar radiation was artificially changed. In Period I the incident radiation was increased by 25% from the observed value, while in Period II the incident radiation was reduced by 25% from the observed values.

The results of Experiment 2 are illustrated in Fig. 8. Figure 8(a) shows the time series of the observed solar radiation incident at the ground for the 92-day period, while Fig. 8(b) gives the artificially changed, incident solar radiation used in the perturbed runs. Figures 8(c) and (d) illustrate the modelled soil moisture in the surface and the deep layer, respectively, and Figs. 8(e) and (f) present the modelled sensible- and latent- heat fluxes, respectively. In this experiment, the modelled soil moistures in the control run and the perturbed run without soil-moisture nudging are fairly similar, while the application of nudging reverses the sign of the discrepancy between the control and that in perturbed runs in both Period I and Period 11. This is because the nudging process is misinterpreting the error in surface variables. For example, in Period I, the increased surface radiation tends to dry the soil (dashed line). However, when nudging is applied, the increased radiation tends to increase the surface temperature and thus the near- surface air temperature. This is misinterpreted as being due to a low soil moisture, and the nudging (wrongly) increases the soil moisture in both the surface and deep layers (full line with circle). The converse is true in Period 11.

Figures 8(e) and (f) show that, with inappropriate incident solar radiation, soil- moisture nudging tends to correct the sensible-heat flux by introducing an error into the modelled latent-heat flux. This is because nudging is more sensitive to changes in the near-surface air temperature than to the (absolute value of) near-surface humidity. For example, in Period I, the increased radiation tends to increase the turbulent fluxes. The predominant increase in near-surface temperature gives a (nudged) increase in soil moisture as described above. There is a downward correction in the sensible-heat flux to bring the modelled air temperature and relative humidity closer to the observed values, but the latent-heat flux is then necessarily miscalculated to preserve the surface energy budget.

In the introduction, mention was made of the fact that the ECMWF model has a dry-warm drift which results from an underestimation of cloud cover with consequent overestimation of solar radiation. In Fig. 8, during Period I, the dashed lines in some respects simulate this dry-warm drift. The figure suggests that certain dry-warm drift symptoms, such as deviations in soil moisture, sensible-heat flux, and temperature (not shown), are eased by the application of soil-moisture nudging. However, if the overestimation of solar radiation is not subsequently corrected by the cloud scheme in the model, some variables, like the latent-heat flux, are adversely affected by nudging.

1898

k (a) ' € 400

Y. HU et nl.

190 200 210 220 230 240

'E

p 400 300

200 ; 100 E

x n 150 160 170 180 190 200 210 220 230 240

5 v--

0.4 I I I I I I I I

e!

0.3 - p 0.2 8 c z 0.1

150 160 170 180 190 200 210 220 230 240

0.4 \ ' I I I I ' I 1 I I

150 160 170 180 190 200 210 220 230 240

m- 0

E

51 -100 6

E

X -50 I 9

1 50 160 110 1 80 190 200 210 220 230 240

150 160 170 180 190 200 210 220 230 240 Julian day (1 987)

Figure 8. Same as Fig. 7, but for Experiment 2 (see text).


There is no active cloud scheme to produce feedback from the soil-moisture nudging in the SCM used in this study, and studies of the effectiveness of soil-moisture nudging in compensating dry-warm drift via cloud-cover modification are needed using a fully coupled version of the model.

In summary, on the basis of this experiment, soil-moisture nudging cannot be expected to compensate for unrealistically simulated surface radiation in an NWP model. Erroneous estimates of soil moisture (worse than those given without nudging) will result. Application of soil-moisture nudging with erroneous radiation will result in improved estimates of the surface sensible-heat flux and near-surface temperature, but the latent-heat flux (and, by implication, the absolute value of near-surface atmospheric humidity) will be miscalculated to satisfy the surface energy balance. The generalization of such results to a global atmospheric model might be difficult, due to the fact that, in a one-dimensional model, feedbacks with the atmosphere through clouds and precipitation are not represented. However, in order to avoid spurious corrections, this study shows the importance of identifying atmospheric situations when errors in near- surface parameters do not come from soil moisture.

4. SUMMARY A N D CONCLUSIONS

The research described in this paper is a critical evaluation of the nudging technique developed to re-initialize (update) soil moisture in NWP models from the knowledge of near-surface parameters and short-range forecast errors. A single-column version of the ECMWF model was used to perform simulations at 16 sites selected to sample a range of climates and land cover across the globe, with forcing taken from the ECMWF analysis at these locations in both the northern hemisphere summer and the northern hemisphere winter. It is important to recognize that use of the identical-twin approach (much used in this study) is known to lead to overly optimistic statements because nature is more complex than the model used in the reference simulation. Moreover, many aspects of the land-surface/atmosphere interaction will influence predicted error in near-surface temperature and humidity. Only some of these are represented in the model we use in this study and this influences the degree to which our results can be regarded as independent of the model used. However, in general, we conclude that soil- moisture nudging can be a useful tool for improving the forecast of near-surface weather variables and surface energy fluxes in an NWP model as shown by Viterbo and Courtier ( 1 995) and Giard et al. (1 996).

The application of a PCA to identify the independent and dependent components of near-surface parameter forecast errors, followed by the derivation of soil-moisture nudging coefficients relative to the first principal component of these two variables, provides a mechanism for defining nudging coefficients that are more stable. However, using this improved procedure cannot, of course, compensate for the fact that the application of soil-moisture nudging assumes that near-surface variable forecast errors are linked to soil-moisture errors, which is likely to be true when the surface-energy exchange is large enough. For this reason, the nudging coefficients are only plausibly derived, and soil-moisture nudging can only plausibly be applied during the (local) summer, except at low-latitude sites where the method is likely to be relevant all year.

The fact that near-surface temperature and relative humidity forecast errors are strongly correlated suggests that using just one of these two variables for soil-moisture nudging might be sufficient. This study showed that nudging with air temperature (but not relative humidity) is possible, and that it is nearly as effective as using the first principal component of temperature and relative humidity.

1900 Y. HU et al.

The original motivation for this study was to seek globally applicable nudging coefficients for the 01 method that were parametrized in terms of the vegetation-related parameters used in the ECMWF model. However, the values of PCA soil-moisture nudging coefficients derived for the northern hemisphere summer are sufficiently similar to propose using their all-site, daily-average values as the basis of a globally applicable soil-moisture scheme. This study showed that using these values provides soil-moisture nudging which is as good as using site-specific, time-specific values at our 16 study sites, thus confirming the value of using these globally applicable averaged values.

Long-term tests of the application of soil-moisture nudging methods (using globally applicable PCA-based nudging coefficients) at the FIFE site gave important insight into issues involved with the use of soil-moisture nudging. First, soil-moisture nudging will provide good estimates of near-surface weather variables and surface fluxes if used to correct for poorly calculated precipitation in an NWP model. However, it does not give an accurate determination of the location of the soil moisture within layers of soil that are modelled as being directly accessible to the atmosphere. Second, if used in an NWP model which has good simulation of precipitation but poorly simulated surface radiation, soil-moisture nudging will give good estimates of the surface sensible-heat flux and near-surface air temperature, but may give poor estimates of soil-moisture status and surface latent-heat flux. This result shows that some additional information is required to identify the source of atmospheric errors before the nudging technique can be successfully applied.

A more detailed evaluation of soil-moisture nudging methods using globally applicable soil-moisture nudging coefficients derived using PCA methods is required in- volving long-term tests with the technique implemented in a three-dimensional weather forecast model. In particular, the PCA approach, which uses only one set of global coefficients obtained from the knowledge of forecast error statistics, should be compared with the empirical approach developed by Viterbo (1996). In the near future, most operational data-assimilation systems will use a variational approach instead of an 01. The four-dimensional variational assimilation is very appealing for the analysis of soil moisture because it can use observations that are linked to model variables through non- linear relationships. This method also accounts explicitly for the temporal distribution of observations. Encouraging results have already been obtained by Mahfouf (1 99 1) and more recently by Callies et al. (1999) using single-column models. Over continents, satellite radiances contain information about soil moisture that could be used in data- assimilation systems. Van Den Hurk et al. (1998) have demonstrated the usefulness of infrared skin temperature to retrieve soil moisture in semi-arid regions. Calvet et al. (1998) have shown that soil moisture in the root-zone can be inferred over vegetated areas from the knowledge of surface soil moisture. This feasibility study is of interest for future use of microwave satellite data, because microwave signals only penetrate the first centimetres of soil.

ACKNOWLEDGEMENTS

Primary support for this research was provided under the Global Energy and Water Cycle Experiment Continental-scale International Project through a National Oceano- graphic and Atmospheric Administration Grant NA46GPK0247. Two of the authors (X. Gao and H. Gupta) were supported under a National Aeronautics and Space Admin- istration Grant NAGW-2425.


Andre, J. C., Goutorbe, J.-P., Pemer, A,, Becker, F., Bessemoulin, P., Bougealt, P., Brunet, Y., Brutsaert, W., Carlson, T., Cuenca, R., Gash, J . , Gelpe, J., Hildebrand, P., Lascouarde, J. P., Lloyd, C., Mahrt, L., Mascart, P., Mazaudier, C., Noilhan, J., Ottle, C., Payen, M., Phulpin, T., Stull, R., Shuttleworth, J., Schmugge, T., Taconet, O., Tarrieu, C., Thepenier, R.-M., Valencogne, C., Vidal-Madjar, D. and Weill, A.

Miller, M. J. and Betts, A. K. Beljaars, A. C. M., Viterbo, P.,

Betts, A. K., Viterho, P. and Beljaars, A. C. M.

Bouttier, F., Mahfouf, J.-F. and Noilhan, J.

Callies, U., Rhodin, A. and

Calvet, J.-C., Noilhan, J. and Eppel, D. P.

Bessemoulin, P.

Dickinson, R. E., Henderson-Sellers, A., Kennedy, P. J. and Wilson, M. F.

Gao, X., Sorooshian, S. and Gupta, H. V.

Garratt, J. R.

Giard, D., Bazile, E., Noilhan, J. and Douville, H.

Houser, P. R., Shuttleworth, W. J . , Gupta, H. V., Famiglietti, J. S., Syed, K. H. and Goodrich, D. C.

Bastiaanssen, W. G. M., Pelgrum, H. and Meijgaard, E. V.

Jin, M., Dickinson, R. E. and Vogelmann, A. M.

Hurk, V. D.,

Lorenc, A. C.

Mahfouf, J. F.

Manabe. S.

REFERENCES

1988 HAPEX-MOBILHY: First results from the special observing period. Annales Geophysica, 6(5), 477492

1996

1998

1993a

1993h

1999

1998

1986

1996

1993

1996

1998

1997

1997

1986

1991

1969

The anomalous rainfall over the USA during July 1993: sensitivity to land surface parameterization and soil moisture anoma- lies. Mon. Weather Rev., 124(3), 362-383

Comparison of the land-surface interaction in the ECMWF re- analysis model with the 1987 FIFE data. Mon. Weather Rev., 126,186198

Sequential assimilation of soil moisture from atmospheric low- level parameters. Part 1: Sensitivity and calibration studies. J. Appl. Meteorol., 32, 1335-1351

Sequential assimilation of soil moisture from atmospheric low- level parameters. Part 2: Implementation in a mesoscale model. J. Appl. Meteorol., 32, 1352-1 364

A case study on variational soil humidity analysis from atmospheric observations. J. Hydrol., in press

Retrieving the root-zone soil moisture from surface soil moisture or temperature estimates: a feasibility study based on field measurements. J. Appl. Mereorol., 37,37 1-386

‘Biosphere-Atmosphere Transfer Scheme (BATS) for the NCAR community model’. NCAR Technical Note, NCAR/TN- 275+STR

Sensitivity analysis of the Biosphere-Atmosphere Transfer Scheme. J. Geophys. Res., 101(3), 7279-7289

Sensitivity of climate simulations to land-surface and atmospheric boundary layer treatments-a review. J. Climate, 6,419449

‘Assimilation of soil water content and temperatures in a global NWP model’. Pp. 205-206 in Proceedings of the 2nd international scientific conference on the global energy and water cycle, 17-21 June 1996, Washington D.C. International GEWEX Project Office, Washington, D.C.

Integration of soil moisture remote sensing and hydrological mod- eling using data assimilation. Water Resoul: Res., 34(12), 3405-3420

A new methodology for assimilation of initial soil moisture fields in weather prediction models using Meteosat and NOAA data. J. Appl. Meteorol., 36, 1271-1283

A comparison of CCM2BATS skin temperature and surface-air temperature with satellite and surface observations. J. Cli- mate, 10,1505-1523

Analysis methods for numerical weather prediction. Q. J. R. Meteorol. Soc., 112, 1177-1 194

Analysis of soil moisture from near-surface parameters: a feasibility study. J. Appl. Meteorol., 30, 1534-1547

Climate of the ocean circulation. Part I. Mon. Weather Rev., 97, 739-714

1902 Y. HU et al.

Milly, P. C. D. and Dunne, K. A.

Mintz, Y. and Walker, G. K.

1994

1993

Oglesby, R. J. 1991

Rowntree, P. R. and Bolton, J. A. 1983

Shukla. J. 1984

Viterbo, P. 1996

Viterbo, P. and Beljaars, A. C. M. 1995

Viterbo, P. and Courtier, P. 1995

Sensitivity of the global water cycle to the water-holding capacity of land. J. Climate, 7,506-526

Global fields of soil moisture and land surface evapotranspiration derived from observed precipitation and surface air temperature. J. Appl. Meteorol., 32, 1305-1 334

Spring time soil moisture, natural climatic variability, and North American drought as simulated by the NCAR community climate model. J. Climate, 4, 890-897

Simulation of the atmospheric response to soil moisture anoma- lies over Europe. Q. J. R . Mefeorol. Soc., 109,501-526

Predictability of time averages. Part 11: The influence of the boundary forcing. Pp. 155-206 in Problems andprospects in long and medium range weatherforecaszing. Eds. D. M. Bur- ridge and E. Kallen. Springer-Verlag

‘The representation of surface processes in general circulation models’. Ph.D. Thesis, Lisbon

An improved land surface parameterization scheme in the ECMWF model and its validation. J. Climate, 8,2716-2748

‘The importance of soil water for medium-range weather forecasting, implications for data assimilation’. Pp. 121-130 in Pro- ceedings of Workshop on imbalances of slowly varying components of predictable atmospheric motions, March 1995, Beijing, China. WMO

Soil-moisture nudging experiments with a single-column version of the ECMWF model

Documents