Improving runoff prediction through the assimilation of the ASCAT soil moisture product

The role and the importance of soil moisture for meteorological, agricultural and hydrological applications is widely known. Remote sensing offers the unique capability to monitor soil moisture over large areas (catchment scale) with, nowadays, a temporal resolution suitable for hydrological purposes. However, the accuracy of the remotely sensed soil moisture estimates has to be carefully checked. The validation of these estimates with in-situ measurements is not straightforward due the well-known problems related to the spatial mismatch and the measurement accuracy. The analysis of the effects deriving from assimilating remotely sensed soil moisture data into hydrological or meteorological models could represent a more valuable method to test their reliability. In particular, the assimilation of satellite-derived soil moisture estimates into rainfall-runoff models at different scales and over different regions represents an important scientific and operational issue. In this study, the soil wetness index (SWI) product derived from the Advanced SCATterometer (ASCAT) sensor onboard of the Metop satellite was tested. The SWI was firstly compared with the soil moisture temporal pattern derived from a continuous rainfall-runoff model (MISDc) to assess its relationship with modeled data. Then, by using a simple data assimilation technique, the linearly rescaled SWI that matches the range of variability of modelled data (denoted as SWI ) was assimilated into MISDc and the model performance on flood estimation was analyzed. Moreover, three synthetic experiments considering errors on rainfall, model parameters and initial soil wetness conditions were carried out. These experiments allowed to further investiCorrespondence to: L. Brocca (l.brocca@irpi.cnr.it) gate the SWI potential when uncertain conditions take place. The most significant flood events, which occurred in the period 2000–2009 on five subcatchments of the Upper Tiber River in central Italy, ranging in extension between 100 and 650 km2, were used as case studies. Results reveal that the SWI derived from the ASCAT sensor can be conveniently adopted to improve runoff prediction in the study area, mainly if the initial soil wetness conditions are unknown.

sensed soil moisture estimates has to be carefully checked. The validation of these estimates with in-situ measurements is not straightforward due the well-known problems related to the spatial mismatch and the measurement accuracy. The analysis of the effects deriving from assimilating remotely sensed soil moisture data into hydrological or meteorological models could represent a more valuable method to test their reliability. In particular, the assimilation of satellite-derived soil moisture estimates into rainfall-runoff models at different scales and over different regions represents an important scientific and operational issue.
In this study, the soil wetness index (SWI) product derived from the Advanced SCATterometer (ASCAT) sensor onboard of the Metop satellite was tested. The SWI was firstly compared with the soil moisture temporal pattern derived from a continuous rainfall-runoff model (MISDc) to assess its relationship with modeled data. Then, by using a simple data assimilation technique, the linearly rescaled SWI that matches the range of variability of modelled data (denoted as SWI * ) was assimilated into MISDc and the model performance on flood estimation was analyzed. Moreover, three syn- 20 thetic experiments considering errors on rainfall, model parameters and initial soil wetness conditions were carried out. These experiments allowed to further investigate the SWI potential when uncertain conditions take place. The most significant flood events, which occurred in the period 2000-2009 on five subcatchments of the Upper Tiber River in Central Italy, ranging in extension between 100 and 650 km 2 , were used 25 as case studies. Results reveal that the SWI derived from the ASCAT sensor can be conveniently adopted to improve runoff prediction in the study area, mainly if the initial soil wetness conditions are unknown.

1 Introduction
Soil moisture plays a fundamental role in the partitioning of rainfall into runoff and infiltration inside a catchment. In particular, for a given storm event, different values of initial soil moisture conditions can discriminate between minor or catastrophic flooding effects (see e.g., Crow et al., 2005;Brocca et al., 2008;Berthet et al., 2009;Merz and 5 Bloschl, 2009). Therefore, the assimilation of soil moisture information within rainfallrunoff models can provide, in theory, a great improvement for both runoff prediction and forecasting.
Several studies investigated the use of soil moisture observations within rainfallrunoff models by using, basically, three different methodologies. Some authors di-10 rectly used these observations for the assessment of the antecedent wetness conditions through in-situ (Meyles et al., 2003;Longobardi et al., 2003;Brocca et al., 2009b;Huang et al., 2007;Pfister et al., 2003;Tramblay et al., 2010;Zehe et al., 2010) and remotely sensed (Jacobs et al., 2003;Goodrich et al., 1994;Brocca et al., 2009a;Beck et al., 2010) estimates. For instance, some studies (Brocca et al., 2009b;Tramblay et 15 al., 2010;Beck et al., 2010) investigated the relationship between soil moisture measurements (in-situ and from satellite sensors) and the soil potential maximum retention parameter of the Soil Conservation Service -Curve Number method (Chow et al., 1988) that can be considered a good indicator of the wetness conditions at the catchment scale. Other applications employed soil moisture data for the improvement of the 20 calibration and testing of rainfall-runoff models (Parajka et al., 2006(Parajka et al., , 2009Wooldridge et al., 2003;Koren et al., 2008) obtaining that, only in certain conditions, the insertion of soil moisture information can help the model structure identification and parameterization. Finally, sequential data assimilation techniques with in-situ (Loumagne et al., 2001;Aubert et al., 2003;Anctil et al., 2008) Aubert et al., 2003). The explanation of the little improvements so far obtained by assimilating soil moisture observations can mainly be related to three aspects (see also Vereecken et al., 2008;Crow and Ryu, 2009): 1. the spatial mismatch: the measurement extent is, usually, too low for in-situ data and too coarse for satellite sensors when compared with model quantities; 2. the limited soil moisture data availability: only recently in-situ soil moisture networks are being set up and satellite data with daily, or lower temporal resolution (required for hydrological applications) are becoming available; 3. the layer depth: remote sensing sensors are able to investigate only a thin surface layer (2-5 cm) that does not match the soil depth (1-2 m) usually simulated within rainfall-runoff models.
Notwithstanding these drawbacks, nowadays remote sensing represents the best possible choice for soil moisture monitoring over large areas. In fact, the first satellite mis-  2007(Bartalis et al., 2007 with coarse spatial resolution (25-50 km) and nearly daily time scale. Indeed, Sinclair and Pegram (2010), for the South Africa region, compared the soil moisture simulations produced by the rainfall-runoff TOPKAPI (TOPographic Kinematic APproximation and Integration) model to the exponentially filtered time series of 5 the ASCAT surface soil moisture obtaining a good linear agreement in the dynamic behavior of the two independent soil moisture estimates. The authors inferred that the soil moisture fields obtained by the combination of ASCAT satellite data and land surface modeling (TOPKAPI) will be valuable for flash flood guidance and other applications in the region.

10
Based on the above insights, the main purpose of this study is the assessment of the effects that can be obtained by assimilating the ASCAT soil moisture product in rainfallrunoff modelling. In particular, to our knowledge, this is the first study where coarse resolution satellite data are assimilated into a rainfall-runoff model by using a real data set. In fact, previous studies with real data sets (Pauwels et al., 2001(Pauwels et al., , 2002Francois et al., 2003;Matgen et al., 2006) considered the assimilation of high resolution SAR (Synthetic Aperture Radar) images that, however, suffer of the well-known problem of the low revisit time (∼30 days). On the other hand, studies using coarse resolution sensors employed satellite data only for rainfall-runoff model calibration (Parajka et al., 2006(Parajka et al., , 2009 or for the assessment of the reliability of the modeled soil mois-20 ture (Sinclair and Pegram, 2010). Moreover, unlike previous studies, the assessment of the benefits (if any) of this assimilation was carried out considering an hourly time scale, which is the appropriate time resolution for operational flood forecasting activities over small to medium catchment sizes (<5000 km 2 ). Specifically, five subcatch- assimilation of ASCAT when errors on rainfall, model parameters or initial conditions take place.

Methods
In the following, the rainfall-runoff model, the derivation of the satellite soil moisture index and the data assimilation technique are described. The soil moisture product 5 derived from ASCAT is represented by a dimensionless index between 0 and 1 and, hence, to be consistent throughout the paper, for it the more appropriate term "saturation degree" is used hereafter, instead of "soil moisture".

Rainfall-runoff model
The continuous rainfall-runoff model, MISDc ("Modello Idrologico SemiDistribuito in The model consists of two components; the first one is a soil water balance model that simulates the saturation degree temporal pattern and sets the initial conditions of an event-based rainfall-runoff model for flood hydrograph simulation (second component). 15 By coupling the two models through an experimentally derived relationship (Brocca et al., 2009b), the structure of a parsimonious and robust continuous rainfall-runoff model is derived. The model requires, as input data, meteorological variables routinely measured (rainfall and air temperature) and incorporates a limited number of parameters. In particular, a total of 6 parameters have to be estimated in the model, 4 for the soil water 20 balance model and the remaining 2 for the event-based rainfall-runoff model. MISDc furnishes, as output, the discharge at the catchment outlet and the saturation degree (SD), at the catchment scale for the soil layer influencing the rainfall-runoff transformation, denoted henceforth as "modeled saturation degree", SD mod (t

Soil wetness index
The assimilation of remotely sensed saturation degree estimates, representative of a soil layer depth of 2 to 5 cm, into rainfall-runoff models, simulating a layer depth usually greater than 1 m, can be conducted in two different ways: (i) direct insertion of the surface satellite data into the model that, however, should incorporate a surface 5 layer ad hoc implemented for the assimilation (Parajka et al., 2009); (ii) assimilation of a profile soil wetness product obtained by the surface one through the application of an appropriate technique (Parajka et al., 2006;Sinclair and Pegram, 2010). In this study, this latter option was selected and the semi-empirical approach (also known as exponential filter) proposed by Wagner et al. (1999a) was adopted for this purpose 10 both for its simplicity and for the request of a single parameter. Briefly, this method assumes that the variation of the average profile saturation degree is linearly related to the difference between the surface and profile saturation degree. For that, the profile saturation degree can be obtained by the knowledge of the surface saturation degree and a parameter T , named characteristic time length, representing the time scale of where m s (t n ) is the surface saturation degree observed by the satellite sensor, SWI(t n ) is the Soil Wetness Index representing the profile averaged saturation degree and t n is 20 the acquisition time of m s (t n ). The gain K n at time t n is given by (in a recursive form): and it ranges between 0 and 1. For the initialization of this filter, K 1 and SWI 1 were set to 1 and m s (t 1 ), respectively (Albergel et al., 2009 Prior to the assimilation of SWI(t) into the rainfall-runoff model, a simple linear normalization was conducted through the following equation: where t is time, SWI * (t) is the linearly rescaled SWI(t), the bar¯and σ are the mean and standard deviation operators and SD mod (t) is the modeled saturation degree. The basis 5 of Eq. (3) is to surmise that SWI * (t) and SD mod (t) have the same mean and standard deviation. This transformation is required because the exponential filter, Eq. (1), mainly for large T values, has the effect of smoothing the satellite-based surface saturation degree time series. Consequently, the variability range of SWI is no more limited by 0 and 1 as the saturation degree but varies in a narrower range. Moreover, to assimilate the satellite data into the MISDc model, the range of variability of SD mod and of the ASCAT data has to be the same (Aubert et al., 2003).

Data assimilation technique
In this preliminary study, a very simple approach is employed for the assimilation of the remotely sensed soil wetness index into MISDc model. A nudging scheme was used to fact, G can be written as (Maybeck, 1979): where σ 2 SD mod and σ 2 SWI * are the model and the SWI * error variance in the retrieval of saturation degree, respectively. For G=1 the observations are assumed right, σ 2 SWI * =0, (direct insertion), vice versa for G=0 the model is assumed perfect. Although the nudg-5 ing scheme employed in this study is not optimal in a statistical sense, it is a computationally inexpensive approach to address the issue on the usefulness of the ASCAT observations to improve runoff prediction (Scipal et al., 2008). In the remainder, for sake of simplicity, the time dependence of SD mod (t) and SWI * (t) is omitted.

Study area and data set 10
The study area is the Upper Tiber River catchment that is located in an inland region of Central Italy. In particular, five subcatchments were used to test the assimilation of the ASCAT saturation degree product: Niccone at Migianella (NIC, 137 km 2 ), Assino at Serrapartucci (ASS, 165 km 2 ), Chiani at Ponticelli (CHI, 100 km 2 ), Tevere at S. Lucia (TEV, 658 km 2 ) and Timia at Cantalupo (TIM, 549 km 2 ). Figure 1 shows the frame- 15 work of the study area along with the available hydrometeorological network and the centroids of the ASCAT time series. Table 1 reports the main characteristics of each catchment in terms of topography, soil type, land use and hydrometeorological network density.
The climate is Mediterranean with mean annual rainfall of ∼950 mm and ranging over 20 the region between 650 mm and 1600 mm (based on the period 1951-1999 and on a network of more than 60 raingauges). Higher monthly rainfall values generally occur during the autumn-winter period when floods occur. Mean annual air temperature (for the same period 1951-1999) ranges between 3.5 • C and 14.0 • C, with maximum in July and minimum in January. Accordingly, the mean annual potential evapotranspiration computed with the Thornthwaite formula is, on average, about 800 mm. In the study area, a dense hydrometeorological monitoring network (1 station every 150 km 2 ) has been operating for more than 25 years and the data are recorded with a time interval of 30 min. The existing network consists of 84 rain gauges, 36 ther-5 mometers and 43 hydrometric gauges. For this study 10 years of rainfall and temperature data recorded from January 2000 to December 2009 were employed.

ASCAT data set
Succeeding the ERS-1 and ERS-2 scatterometers, ASCAT is a real-aperture radar instrument measuring radar backscatter at C-band in VV polarization. Because ASCAT 10 operates with two sub-swaths, more than twice of the ERS-1/2 scatterometer coverage is provided. For Western Europe, measurements are generally obtained twice a day, one in the morning (descending orbit) and one in the evening (ascending orbit), between 08:00-11:00 and 17:00-21:00 UTC, respectively. Soil moisture is retrieved from the ASCAT backscatter measurements using a time series-based change de-15 tection approach previously developed for the ERS-1/2 scatterometer by Wagner et al. (1999a,b,c). In this approach soil moisture is considered to have a linear relationship to backscatter in the decibel space, while the noise sources include the instrument noise, speckle and azimuthal anisotropies. The surface roughness is assumed to have a constant contribution in time, and therefore is not accounted for in the change detec-20 tion algorithm. By knowing the typical yearly vegetation cycle and how it influences the backscatter-incidence angle relationship for each location on the Earth, the vegetation effects can be removed (Wagner et al. 1999b), revealing the soil moisture variations. As a last step, the historically lowest and highest values of observed soil moisture are assigned to the 0% (dry) and 100% (wet) references, respectively, thereby yielding time Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | measurements using the algorithm described by Naeimi et al. (2009). First validation studies of the ASCAT soil moisture product assessed its reliability for estimating both in-situ and ground modeled soil moisture observations across different region in Europe (Albergel et al., 2009;Brocca et al., 2010b,d) and South Africa (Sinclair and Pregram, 2010) both considering m s and SWI.

4 Overall methodology and synthetic data assimilation experiments
For each catchment, our overall approach is based on the application of the MISDc model to simulate the most significant flood events occurred before the period considered for the assimilation of the ASCAT SWI data (i.e. the period [2000][2001][2002][2003][2004][2005][2006][2007]. Then the model is run for the period for which ASCAT data are available (January 2008-Juny 10 2009) and SD mod is compared with the SWI. Afterwards, the SWI is rescaled to match the range of variability of the modeled data through Eq. (3). Thus, the rescaled SWI, SWI * , is assimilated in the MISDc and the outcomes on flood prediction are compared with those without assimilation. In addition, three synthetic experiments are conducted as follows. Firstly, the rain- 15 fall temporal pattern is perturbed through a multiplicative scaling factor sampled from a mean-one log-normal distribution with a dimensionless standard deviation of 0.4 (Crow and Ryu, 2009). Moreover, a possible bias in the rainfall pattern is also analyzed considering a log-normal distribution with the mean value greater than one. Secondly, an error in the model parameters was simulated with a multiplicative scal-20 ing factor sampled from a mean-one normal distribution with a dimensionless standard deviation of 0.4. Since within MISDc the soil water balance model computes the saturation degree temporal pattern, an error only in its parameters was also considered. In fact, these parameters are those more variable across catchments and, hence, they are more difficult to be estimated. In both cases, and also in the normal data assimi- conditions. Thirdly, an error in the initial conditions was tackled on starting the model run one month before the occurrence of the first flood event and considering as initial soil wetness condition the whole range of feasible values from zero to one. For all cases, as previously, the model results with and without assimilation were compared and discussed.

5
To evaluate the efficiency of the assimilation procedure, an efficiency index (Eff), was used (Aubert et al., 2003): where t is time, Q obs is the observed discharge, Q sim and Q ass are the simulated discharge without and with assimilation, respectively. If Eff is greater than 0 then the 10 assimilation produces an improvement in the runoff simulation by the model.

Results and discussion
In this section we show the MISDc model calibration, the comparison between the modeled and ASCAT-derived saturation degree and finally the data assimilation results. 15 The rainfall-runoff model MISDc was calibrated for the five investigated catchments considering more than 100 flood events that occurred in the period 2000-2007. In fact, it has to be noticed that the model is able to simulate only flood events (not the continuous discharge time series) because it was developed for this purpose in order to reduce the number of involved parameters as much as possible. Therefore, all the and on runoff volume, ε Rd , and the Nash-Sutcliffe efficiency index (NS). The model results (not shown for sake of brevity) were found to be in good agreement with observed data with median NS higher than 75% and median absolute error on peak discharge, |ε Qp |, and runoff volume, |ε Rd |, less than 30%. Then, the model was run for the period for which ASCAT data are available. This period can consequently be considered as 5 a validation period. Obviously, a limited number of flood events were available for this short period (∼4 events for each catchment) and, hence, the obtained results have to be considered as preliminary. These results are reported in Table 2 showing a slightly lower accuracy than that obtained during the calibration period (as expected). This is mainly due to the limited magnitude of the flood events observed in the last 2 years and 10 half. More detailed information on the model behaviour and performance in the study area can be also found in Brocca et al. (2010a,e).

Modeled versus ASCAT-derived saturation degree
Based on the model results for the period 2008-2009, the saturation degree simulated by the model, SD mod , was used for the assessment of the ASCAT SWI reliability. In 15 particular, for each catchment, the corresponding SD mod was compared with the SWI index derived from the ASCAT pixel closest to the centroids of the catchment itself. The comparison between the SWI and SD mod is shown in Fig. 2 for four of the investigated catchments. As it can be seen, there is a very good agreement between the temporal pattern of the two quantities, with a determination coefficient, R 2 , higher than 0.90 20 with a maximum of 0.95 for the NIC catchment. Moreover, the Root Mean Square Error computed between SD mod and the rescaled SWI, SWI * , was less than 0.046.
Assuming for the soil a porosity equal to 0.30, this corresponds to an error of less than 0.014 m 3 /m 3 in volumetric terms.  Pellarin et al., 2006;De Lange et al., 2008;Brocca et al., 2010b). As it can be inferred from Eq. (1), the computation of SWI requires the calibration of the T parameter, which was carried out through the comparison with SD mod . According to previous studies (Brocca et al., 2009a), T was found ranging between 30 (TEV) and 90 days (TIM). In a physically consistent way, lowest values were obtained for the 5 fast reacting catchments characterized by a thin active soil layer. Accordingly, the soil layer depth used in the rainfall-runoff model was equal to 70, 80, 120, and 270 cm for the TEV, ASS, NIC and TIM catchments, respectively. However, when considering a constant T value of 50 days, R 2 remained higher than 0.83 for all catchments and so, this value might be adequate for ungauged catchments in the study area.

Data assimilation
The rescaled SWI, SWI * , derived through Eq. (3), was directly incorporated into the MISDc by using Eq. (4). For this step, the G parameter was made to vary between 0 (perfect model) and 1 (perfect observation). Figure 3 shows the value of the Nash-Sutcliffe efficiency index (NS), versus the G parameter for all the investigated catch-15 ments. Overall, it is evident that the assimilation of remotely sensed saturation degree estimates can only improve the model performance as can be already inferred for G>0.2 for NIC, SER and CHI catchments; whereas for TEV and TIM the assimilation did not make worse the model performance. However, by inspecting in depth Fig. 3, an optimal G value (i.e. maximizing the NS index) can be found, on average, equal to 20 0.2; while ranging between 0.01 (TEV) and 0.40 (CHI), with higher values when the added-value of the saturation degree observations was more substantial. According to Eq. (5), a gain parameter equal to 0.2 means that the model error, σ mod , is half of the satellite one. Similar results were obtained by attempting to reproduce saturation degree observations with a satellite and soil water balance model ( The results of the data assimilation procedure in terms of model performance, using the optimal T value for each catchment, are shown in Table 2; Fig. 4 shows, by way of example, the simulated saturation degree for the NIC and ASS catchments (with and without SWI * assimilation) along with the observed and simulated discharge for the flood events which occurred in the study period. For these two catchments satis-5 factory results were derived with the assimilation; the NS coefficient increased from 75 and 62% to 84 and 76% for the NIC and ASS catchments, respectively. In addition, |ε Rd | was reduced by 50 and 33% for the two catchments, respectively, corroborating the added-value of ASCAT assimilation. In fact, since the assimilation modifies the initial wetness conditions, an improvement on total runoff estimation was expected.
10 Therefore, to better visualize the effects of the assimilation of ASCAT on the model simulation, Fig. 5 shows the cumulated runoff for the observed and the simulated data with and without SWI * assimilation for all investigated catchments. As it can be seen, for the three smaller catchments (<150 km 2 ) the improvements related to SWI * assimilation are evident whereas results are quite similar for the other two catchments (TEV 15 and TIM). The same figures can be derived looking at the values of the efficiency index (Eff), that range between 36 and 44% for the three smallest catchments and are less than 8% for the two largest ones.

Synthetic experiments
In order to better analyze the counterintuitive results previously obtained (better re-20 sults for small catchments assimilating coarse resolution satellite data), three synthetic experiments were conducted. The target is to check if the remotely sensed SWI * assimilation can counterbalance a wrong estimation of saturation degree due to errors on rainfall, model parameters or initial soil wetness conditions. For the first two test, as previously described, a random multiplicative error on rain-25 fall and MISDc parameters was considered. For each catchment, 100 simulations were carried out in order to minimize the effects related to randomisation. For both experiments, the increase in model performance due to the SWI * assimilation was reduced  Table 2). On average, the Eff index decreased from 25.6 to 17.2% and 18.8% when rainfall and model parameter errors were considered, respectively. Similar findings were obtained analyzing the other performance indices. It is clear that the error on rainfall and model parameters affected the results significantly. This fact can explain the lower advantages of SWI * assimilation observed for the two larger catchments,

5
where higher uncertainties on model structure suitability (lumped model parameterization) and rainfall pattern identification arise. However, further and more detailed analyses should be carried out to validate this hypothesis. Figure 6 shows the results for these two synthetic experiments applied to the CHI catchment. The three flood events for which a higher added-value was obtained with SWI * assimilation are highlighted (square box). In particular, these events occurred in the transition period between the dry (summer) and wet (winter) saturation degree conditions for which the SWI * assimilation proved to be more valuable. This aspect has to be emphasized because the differences in runoff response for a given rainfall event can be remarkable, in particular for Mediterranean catchments where the seasonal soil moisture temporal pattern is 15 highly pronounced (see e.g., Latron et al., 2009). Moreover, during intermediate wetness conditions, runoff predictability is strongly reduced (Zehe and Bloschl, 2004) and, hence, the improvement deriving from the assimilation of ASCAT data can be highly significant. It has to be noted that when a bias error was exploited to perturb the rainfall time 20 series (as it can occur when satellite or radar rainfall estimates were adopted), the improvement due to SWI * assimilation is more evident. For instance, considering a multiplicative scaling factor with mean equal to 1.1, the Eff score for the NIC catchment increased up to 66%. This is also because MISDc tends to overestimate discharge for this particular catchment (see Fig. 5). Similar results were obtained also for the 25 other two catchments, where an overestimation of discharge by MISDc was observed (see Table 2). On average, the efficiency index was found equal to 55% for these three catchments (NIC, SER and CHI) and only 4% for the remaining two (even a negative efficiency value was found for TEV catchment). As far as the error in model parameters is concerned, if only those included in the soil water balance model are considered, results become more similar to the ones of the normal data assimilation run (see Table 2) with average Eff equal to 26%. All the results obtained through these two synthetic experiments highlight the importance of knowing what could be the source of error when data assimilation findings have to be evaluated. In fact, to understand the results of 5 the data assimilation even for other rainfall-runoff models, satellite sensors or climatic regions, a clear separation of the different error sources has to be investigated. More of interest from the practical and operational point of view is the third synthetic experiment that assumes uncertainties on the model initial conditions. This situation frequently occurs in practice when a continuous time series of the forcing data (rainfall 10 and temperature) is not available because of, for instance, sensor failures, interruptions in the data transmission or errors in real-time data. As expected, in this case the data assimilation furnished a decisive improvement in model performance for all investigated catchments (see Table 2). For instance, Fig. 5 shows the results obtained for the NIC catchment when SWI * was assimilated. MISDc simulations are very similar 15 to those obtained without uncertain initial conditions. Obviously, these results depend on a series of factors including the time in which the initial conditions are set (summer, winter, transition period) and the length and the climatic conditions of the period between the start of the model run and the first flood event. However, these findings are very encouraging for an efficient use of satellite information within operational systems 20 of real-time flood forecasting which frequently rely on event-based rainfall-runoff models and require a strict and accurate estimation of the soil moisture initial conditions (Berthet et al., 2009).

Conclusions
A first investigation of the potential of the ASCAT derived saturation degree product for 25 hydrological applications was conducted. Two main outcomes can be derived from the performed analyses: 1. SWI, the soil wetness index derived from ASCAT, was found strongly correlated with the simulated saturation degree, with determination coefficients (R 2 ) higher than 0.90 and RMSE values less than 0.014 m 3 /m 3 .
2. The assimilation of the rescaled SWI (denoted SWI * ) into the MISDc rainfall-runoff model furnished an evident improvement of the model performance, especially 5 when initial soil wetness conditions were unknown.
These results support the introduction of the ASCAT-derived saturation degree estimates into an operational system for real-time flood forecasting (Brocca et al., 2010c). However, further and more detailed analysis is needed to confirm these results. A more comprehensive data assimilation technique (e.g. ensemble Kalman filter or particle fil-10 ter), a longer time period and a larger number of catchments have to be used for a more robust test of the proposed procedure. Moreover, a different structure for the rainfall-runoff model should be investigated as well. In particular, the introduction of a thin surface layer might allow to directly assimilate the surface soil moisture values retrieved by using remote sensing, without needing of a method which provides profile 15 soil moisture values (such as the exponential filter used in this study). Introduction

Tables Figures
Back Close

Full Screen / Esc
Printer-friendly Version