1. Introduction
Land surface state predictability has been the focus of increasing scientific interest over the past few decades, mainly because of the significant role and complex nature of land–atmosphere interactions that greatly affect weather and climate. Key land surface variables, such as soil moisture and soil temperature, greatly influence the atmospheric boundary layer at both short-range and seasonal time scales (e.g., Beljaars et al. 1996; Betts 2004; Fischer et al. 2007). Accurate prediction of land surface states can be advantageous to regional weather and seasonal forecasts that notably depend on initial and boundary land surface conditions (e.g., Koster and Suarez 2001; Drusch and Viterbo 2007; Papadopoulos et al. 2008). The level of this accuracy is yet not well determined and pertains to the variety of sources and methods used to define the land surface conditions. Land surface models (LSMs) coupled, or not, to atmospheric models, and forced by in situ or remotely sensed hydrometeorological data, constitute the main tool today for the prediction of land surface parameters, suitable for use in the initialization and boundary conditions updating of advanced weather and regional climate models (e.g., Robock et al. 2003; Koster et al. 2004a; Rodell et al. 2005). Thus, the verification and uncertainty characterization in these Land Data Assimilation Systems (LDASs) becomes of increasing importance, especially under the consideration of the latest advances in remote sensing of near-surface soil moisture and other important land surface variables (Kerr et al. 2001; Entekhabi et al. 2004; Walker and Houser 2004; Reichle et al. 2004, 2007).
Direct observations of land surface variables have been generally based on point measurements of soil moisture and energy fluxes from in situ stations, which are limited to a few locations around the globe, either under the auspices of specific field experiments or—less frequently—as part of long-term in situ networks (Robock et al. 2000). Moreover, the fact that efforts to adequately capture soil moisture properties from spaceborne microwave (MW) sensors are still under way (e.g., Walker and Houser 2004; Reichle et al. 2004), or face significant technical challenges (Leese et al. 2001), complements the lack of comprehensive global monitoring of soil moisture. Consequently, LDASs have long become the gap-filling tool to providing global estimates of soil moisture and other land surface variables that can be directly assimilated into weather- and climate-scale models for retrospective studies or forecasting applications (e.g., Walker et al. 2003). The North American LDAS (NLDAS; Mitchell et al. 2004; see also Schaake et al. 2004), the Global LDAS (GLDAS; Rodell et al. 2004), and the Land Information System (LIS; Kumar et al. 2006) are currently three major representatives of these hybrid modeling systems, each one utilizing multiple LSMs and multiple sources of satellite- and ground-based hydrometeorological observations to provide optimal fields of land surface states and fluxes in near-real time.
Among all hydrometeorological variables used to force the various LDASs, precipitation receives the greatest attention. Accurate precipitation measurements are critical for the implementation of the LDAS near-real time simulations (Gottschalck et al. 2005). The requirement for global or extensive regional simulation coverage can be facilitated either with the use of global–regional climate model rainfall outputs or the exploitation of advanced satellite rainfall retrievals. Both pathways have advantages and disadvantages, yet the ongoing development of high-resolution (<0.25° and <3 h) global rainfall estimates from a combination of infrared (IR) and passive microwave (PMW) retrievals (e.g., Sorooshian et al. 2000; Joyce et al. 2004; Huffman et al. 2007) has driven attention to remotely sensed precipitation estimation. Such datasets have been recently under thorough investigation and intercomparison to define the optimal rainfall estimates based on regional- and time-scale criteria (e.g., Ebert et al. 2007; Tian et al. 2007; Hossain and Huffman 2008; Anagnostou et al. 2010). The anticipated Global Precipitation Measurement (GPM) mission (Smith et al. 2007) is designed to facilitate this effort of providing high-resolution global rainfall estimates from a deployed constellation of satellite-based passive microwave sensors. GPM-era precipitation observations are expected to enhance the accuracy and spatiotemporal resolution of the precipitation forcing in LDASs, leading to improvements in the prediction of land surface states. With the upcoming GPM mission being assessed in terms of societal values, we now need to identify and prioritize hydrologic uses of the anticipated data sources and methods to make this mission as effective as possible over land.
Remote sensing of rainfall from satellite data sources is subject to error that can be of complex structure at high spatiotemporal scale (Hossain and Anagnostou 2004, 2006). The propagation of this error through the nonlinear land–atmosphere interaction processes resolved by LSM can affect soil moisture prediction in a way that depends on scale, precipitation error characteristics, and the complexity of modeling system. This stresses the need for thorough and systematic investigations and quantification of the error propagation properties for optimal LDAS simulations. A key feature for the accurate definition of soil moisture prediction uncertainty is the quantification of interactions between the rainfall forcing uncertainty with the LSM parametric error (Hossain and Anagnostou 2005a,b). Beyond the aforementioned studies and the study by Gottschalck et al. (2005), this subject has not received proper attention in recent literature. Gottschalck et al. (2005) compared multiple satellite-, model-, and ground-based rainfall datasets, which were then used to force year-long GLDAS simulations to perform qualitative diagnosis of their effects on land surface states. They found that GLDAS land surface states are sensitive to different precipitation forcing; percent differences in volumetric soil water content (SWC) between simulations ranged from 75% to 100% for both summer and winter seasons, and these differences were generally 25%–75% less than the percent precipitation differences, indicating that GLDAS, and specifically the Mosaic LSM, acted to generally “damp” precipitation differences. However, there was also evidence of areas where the changes in SWC were equivalent to the precipitation changes. Hossain and Anagnostou (2005a) were the first to explore the issue of the complex interaction between rainfall and modeling uncertainties in LSM soil moisture prediction. They performed numerical experiments using ensemble-based techniques to isolate and characterize the propagation of errors in the satellite rainfall estimation alone, the LSM parametric uncertainty alone (manifesting as nonuniqueness in soil hydraulic parameters), and the combined data–modeling uncertainty. They found that the contribution of precipitation error was generally lower than that of modeling uncertainty, with satellite retrieval error contributing between 20% and 60% of the total uncertainty in soil moisture prediction in the cases of modeling accuracy ranging from low to high, respectively.
In this study, we expand the work by Hossain and Anagnostou (2005a), seeking to examine and quantify the two aforementioned sources of uncertainty in the simulation of soil moisture fields from an offline LSM (member of the GLDAS and LIS) forced by three different sources of remotely sensed precipitation estimates (two from satellites and one from ground radar). The assessment of both sources of error in soil moisture prediction is uniquely facilitated through the use of in situ measurements of soil moisture, rainfall, and other meteorological variables on a small domain in the midwestern United States capturing the state of Oklahoma. The region is covered by a dense network of environmental monitoring stations named the Oklahoma Mesoscale Network (Mesonet; Brock et al. 1995), available over a long-term period (1997–2006). The abundance of measurements over the Mesonet region in combination with the climatic characteristics of this area (e.g., standing in the transition zone between wet and dry areas in the United States; Koster et al. 2004b) made this region suitable for pursuing the goals of our study. The Mesonet station data are used here as reference for evaluating the remotely sensed rainfall retrievals at high spatiotemporal scales as well as the LSM simulations of soil moisture. The study also presents a rigorous benchmarking of the Mesonet network as to its accuracy in deriving area rainfall estimates at the resolution of satellite products (0.25°, 3 h) based on comparisons against the most definitive Micronet station measurements (see section 2 for more details on Mesonet and Micronet). In section 2, we present the area and datasets used in the current study. Section 3 includes a detailed description of the LSM used and the experiments performed. Results on the comparison between Mesonet and Micronet rainfall estimates, rainfall and modeling uncertainty characterization, and the effect of scale on error propagation are presented in section 4. In section 5 we discuss the implications of our study and conclude with the major findings and future directions of this study.
2. Study area and datasets
The current error propagation study was facilitated with the use of data from various sources. First, surface meteorological observations (e.g., pressure, air temperature, relative humidity, solar radiation, wind speed, and rainfall) at 5-min resolution as well as soil moisture observations at three depths (5, 25, and 60 cm) at 30-min resolution were provided from 115 stations of the Mesonet network (Fig. 1). Within the Mesonet network, the U.S. Department of Agriculture (USDA) Agricultural Research Service’s (ARS) Grazinglands Research Laboratory (GRL) has established a smaller area yet denser network of hydrometeorological stations (42 stations from 1994 to 2005, 20 core stations thereafter) that monitors the environmental conditions of the Little Washita watershed, called the Little Washita Micronet. Three stations in the Oklahoma Mesonet are located in the northeast, south, and west areas of the watershed (Fig. 1). Both datasets are quality controlled and flagged for bad quality data. Data are available for the period 1997–2006 for the Mesonet stations and for the period 2002–04 for the Micronet network.
Further, three different remotely sensed rainfall datasets were used for the 3-yr period 2004–06: the National Aeronautics and Space Administration (NASA) Goddard Space Flight Center’s (GSFC) Tropical Rainfall Measuring Mission (TRMM) 3B42 product (obtained at 0.25°–3-h spatiotemporal resolution), the National Oceanic and Atmospheric Administration (NOAA) Climate Prediction Center’s (CPC) morphing technique (CMORPH) product (obtained at 8-km, 30-min spatiotemporal resolution), and the stage IV U.S. Weather Surveillance Radar-1988 Doppler (WSR-88D) radar network [Next Generation Weather Radar (NEXRAD)] estimates (obtained at 4-km–1-h spatiotemporal resolution). Hereafter, the three datasets will be referred to as TRMM, CMORPH, and NEXRAD, respectively.
Measurements from satellite passive microwave sensors are the primary source for the precipitation estimates of both CMORPH and TRMM rainfall products. Multiple satellite platforms are used to facilitate maximum coverage and enhanced temporal sampling for both products, including TRMM, the Defense Meteorological Satellite Program (DMSP), NOAA, and the Earth Observing System (EOS) platforms. However, the way that these sensors are intercalibrated to extract each product differs. In 3B42, PMW sensors are intercalibrated to TRMM’s combined precipitation radar (PR) and TRMM Microwave Imager (TMI) retrievals, whereas CMORPH uses TMI and DMSP Special Sensor Microwave Imager (SSM/I) as a calibration reference, with TMI having the highest precedence whenever available (Tian et al. 2007). Both datasets use IR data from geostationary satellites to fill in MW coverage gaps, yet in different ways. For the CMORPH product, the dynamic morphological characteristics (such as shape and intensity) of the precipitation features are morphed at consecutive times between MW sensor samples by performing a time-weighted linear interpolation. This process yields spatially and temporally continuous passive MW rainfall fields that have been guided by IR imagery and yet is independent of any infrared temperature–based inverion to rainfall rate (Joyce et al. 2004). The TRMM 3B42 product, on the other hand, uses MW-calibrated IR precipitation estimates directly, to fill the MW coverage gaps. A further difference between the two products pertains to the use of surface gauge measurement information: CMORPH only uses satellite estimates, whereas TRMM 3B42 combines the merged MW- and IR-based estimates with gauge observations via scaling of the individual 3B42 3-h precipitation values to monthly gauge analysis. A detailed comparison of these two products in terms of their potential use in LDAS simulations was performed in Tian et al. (2007), who infer that the TRMM 3B42 product is suitable for long-term, retrospective, and climatological studies because of its reduced biases on longer time scales, whereas CMORPH is recommended for short-term applications because of its higher probability of detection of rainfall events. The radar rainfall fields used in this study were extracted from the stage IV National Weather Service (NWS) precipitation estimation algorithm product that involves real-time adjustment of the radar rainfall estimates based on mean-field radar–rain gauge hourly accumulation comparisons and the merging of hourly radar with gauge-interpolated rainfall fields (Fulton et al. 1998; Lin and Mitchell 2005).
The NEXRAD and CMORPH rainfall products were rescaled to the TRMM spatiotemporal resolution (0.25° × 0.25°, 3 h), such that common spatial grid (depicted in Fig. 1) and temporal scale are used for all products to allow for direct comparisons and model input implementation. The in situ Mesonet meteorological data (including the rainfall measurements) were also interpolated to this common spatial grid using the inverse distance weighting (IDW) technique and averaged to 3-h increments. The Kriging interpolation scheme was preferred because of the longer spatial correlation of soil moisture to create the same grid for the Mesonet soil moisture data. A point to note is that both interpolation techniques influence the characteristics of rainfall forcing and soil moisture measurements, mainly in terms of not entirely preserving the rainfall peaks and soil moisture maxima, and thus creating spatially smoother fields. The reader is referred to Anagnostou et al. (2010) for a detailed analysis of the error properties of the aforementioned rainfall products used in this study.
3. Land surface model and experimental setup
The National Center for Atmospheric Research (NCAR) Community Land Model [CLM, version 3.5 (CLM3.5)] is used in this study to simulate the land surface and land–air exchange processes. CLM3.5 is a well-documented model (available online at http://www.cgd.ucar.edu/tss/clm/distribution/clm3.5/index.html) that has been designed to integrate all land processes into a single modeling system. It is one of the models used in GLDAS (Rodell et al. 2004) and LIS (Kumar et al. 2006) and receives extensive attention in current land surface modeling studies (Oleson et al. 2008; Stöckli et al. 2008; Tian et al. 2008).
The model components comprise biogeophysics (i.e., surface fluxes of energy, moisture, and momentum), hydrologic cycle, biogeochemistry, and dynamic vegetation (for the purposes of the current study, we utilized the biogeophysics and hydrology components only). On the basis of externally provided atmospheric forcing data (e.g., precipitation, radiation, wind speed, air temperature, and humidity fields), the model computes a number of prognostic surface variables, including runoff; soil moisture and temperature in various soil layers; water intercepted on the canopy; leaf temperature; and latent and sensible heat fluxes. CLM3.5 has one vegetation layer, like most land surface models, 10 unevenly spaced vertical soil layers (the respective depths are defined at 0.7, 2.8, 6.2, 11.9, 21.2, 36.6, 62.0, 103.8, 172.8, and 286.4 cm) with variable hydraulic conductivity, and up to five snow layers, depending on the total snow depth (Oleson et al. 2004). The land surface is represented by five primary subgrid land cover types (glacier, lake, wetland, urban, and vegetated) for each grid cell. The vegetated portion of a grid cell is further divided into patches of plant functional types (PFTs), each with its own leaf and stem area index and canopy height. Each subgrid land cover type and PFT patch is a separate column for energy and water calculations at every time step. The high-resolution surface datasets currently used in CLM3.5 are based on newly developed Moderate Resolution Imaging Spectroradiometer (MODIS) products (Lawrence and Chase 2007; see Table 1 for the Oklahoma region). Most surface processes—such as evaporation from the ground, transpiration from the plants’ rooting zone, soil and snow water propagation, leaf temperature and fluxes, soil and snow temperature, and phase change—are parameterized through physical equations. The parameterization of runoff-related processes is based on the TOPMODEL concept (Beven and Kirkby 1979; Niu et al. 2005).
Combining CLM3.5 and the datasets described earlier, various experiments were designed. The control experiment included the use of the postprocessed Mesonet meteorological observations as input in CLM3.5 (named CLM–MESONET), while three other experiments were facilitated with the use of the three remotely sensed rain products as rainfall input in place of the respective Mesonet observations (named CLM–TRMM, CLM–CMORPH, and CLM–NEXRAD, respectively). A half-hourly time step was chosen for all CLM3.5 runs to provide for high-accuracy simulations (with a constant rainfall input for each half hour during a 3-h period), while the model outputs were at 3 h to match the satellite rainfall products time sampling. All simulations were 3-yr long, starting at 0000 UTC 1 January 2004 and ending at 0000 UTC 1 January 2007, and were initialized with the CLM3.5 output at 0000 UTC 1 January 2004, produced by an earlier-in-time 7-yr run (1997–2003). This spin-up time is effectively longer in duration than common practice for land surface models (Cosgrove et al. 2003).
Each experiment outputs various soil and energy properties, and our study is focused on the soil moisture fields. It is worth mentioning that a quadratic fitting model was applied to those fields to account for the discrepancy in depth with the actual Mesonet observations. Specifically, soil moisture output at the first four model soil depths (i.e., 0.7, 2.8, 6.2, and 11.9 cm) was taken into account to provide the best available estimate for soil moisture at 5 cm, based on a second-order fitting equation defined uniquely for each grid cell at every time step. The specific design of these four experiments allows for an in-depth analysis of the error propagation of the rainfall products through the soil moisture prediction as well as the relationship between data and modeling uncertainties. The results presented in this study focus on the summer periods [i.e., June–August (JJA)] of 2004 and 2006.
4. Results
a. Assessment of Mesonet sampling on rainfall estimates and soil moisture simulations
The high-resolution Micronet rainfall dataset facilitates an in-depth examination of the coarser-resolution Mesonet network as to its adequacy to provide accurate precipitation forcing data for the land surface model. A Mesonet benchmarking experiment was conducted with CLM3.5 using grid-cell average rainfall from the dense Micronet network as a reference to force a cell in the Mesonet domain shown in Fig. 1 (this is named CLM–MICRONET experiment). Figure 2 shows a comparison between Mesonet and Micronet in terms of the grid-cell average rain accumulations (at 3-h time intervals) and the corresponding model-predicted near-surface (5-cm depth) soil moisture values for the 2004 summer season. As noted from the figure, the Micronet average values of rainfall accumulation are very close to the rainfall accumulation values from the respective Mesonet grid cell, which leads to almost identical temporal evolutions of CLM–MESONET and CLM–MICRONET soil moisture values (see Fig. 2a). The discrepancies observed in Fig. 2a between model results and soil moisture observations by Mesonet stations (i.e., the fact that model results seem to be much more responsive than the observations in terms of response to rainfall events and dry down) is attributed to two main sources: (i) errors because of modeling caused by incorrect model parameters and errors in the assumed-close-to-the-truth reference rainfall and (ii) errors in the soil moisture observations caused by sensor measurement uncertainties and the smoothing effect of the Kriging interpolation scheme applied to the measurements to create the 0.25° grid-cell averages. Unfortunately, the errors in soil moisture observations, which are not quantifiable from currently available data, are lumped into the “modeling” error and do not affect the “rainfall-induced” error. Consequently, significant soil moisture observing errors would skew the relationship between modeling and rainfall-induced land surface model errors; therefore, results presented in this study should be viewed as conservative in terms of the modeling uncertainty.
The discussion that follows offers a quantitative statistical evaluation of the error properties of rainfall products and soil moisture estimates qualitatively captured in Fig. 2a. As expected, the correlation coefficient between rainfall estimates of Micronet and Mesonet is very high (0.95). However, the Mesonet probability of rain detection (POD) relative to Micronet is about 63%. Although this may seem like a low detection score, the majority of the nondetected Micronet rainfall values is below 0.05 mm h−1, as revealed by the conditional POD (at the 0.05 mm h−1 threshold) that is equal to 90%. On the other hand, the false-alarm rate (FAR) is very low in both cases (7% for the unconditional case and 6% for the conditional case). The reader is also urged to look at Anagnostou et al. (2010) for an in-depth statistical analysis of the same Mesonet and Micronet rainfall datasets. Regarding the CLM3.5 soil moisture estimates, the correlation coefficient between CLM–MESONET and CLM–MICRONET is even higher than the respective rainfall value (0.97). Both model estimates are also highly correlated to the Mesonet soil moisture observations (0.88 for CLM–MESONET and 0.84 for CLM–MICRONET), although here the respective bias and error values are quite high because of the spatial interpolation of the initial point measurements and the other error sources discussed in the previous paragraph. The absolute relative errors of these model estimates (defined as absolute relative differences from the respective 2004 summer-season 3-h Mesonet soil moisture observations) are graphically compared in Fig. 2b and seem to exhibit similar characteristics.
Moreover, the relative root-mean-square error (RMSE) of rainfall (defined as the RMS of the difference between Mesonet and Micronet rainfall estimates normalized by the mean value of Micronet rainfall, conditional to Micronet rainfall >0.05 mm h−1) is 0.49 as opposed to a value of 0.07 for the relative RMSE of soil moisture (defined as the RMS of the differences between CLM–MESONET and CLM–MICRONET soil moisture estimates normalized by the mean value of CLM–MICRONET soil moisture). Further, the conditional mean biases are 1.58 for rainfall (Mesonet-to-Micronet rainfall estimates) and 1.02 for soil moisture (CLM–MESONET to CLM–MICRONET soil moisture estimates). These values indicate a “dampening” of the relatively significant rainfall error through the land surface model simulation, verifying similar qualitative (Gottschalck et al. 2005) and quantitative (Hossain and Anagnostou 2005a) results presented in earlier studies. Another aspect that highlights the importance of the latter inference pertains to the dampening of error in rainfall associated with the application of the IDW interpolation technique. The foregoing analysis thus supports our notion that the low-resolution Mesonet observations (as compared to the high-resolution Micronet dataset) are sufficient for studies that include forcing of land surface models and respective error propagation quantification.
b. Rainfall-induced versus modeling-induced errors in soil moisture prediction
Figure 3 depicts the 2004 and 2006 warm-season temporal evolution of observed area-average rain rate and accumulation from the whole Mesonet grid and both observed (Mesonet) and predicted (CLM3.5 experiments) mean near-surface (5-cm depth) soil moisture values. In both warm seasons, we observe a noteworthy correspondence between rain occurrence and changes in both observed and predicted near-surface soil moisture. Enhanced rainfall is always followed by a significant increase in near-surface soil moisture, which is expected but still quite prominent.
Comparing the Mesonet-observed to the model-predicted area-average near-surface soil moisture time series, we notice some differences between the two summer periods under study. In summer 2004, there is good agreement between measured and predicted soil moisture in the study region, especially with regard to CLM–MESONET, CLM–TRMM, and CLM–NEXRAD experiments. However, there are periods when these three experiments seem to overestimate soil moisture as compared to the Mesonet measurements, which coincide with short-term peaks in soil moisture magnitude following events of intense rainfall. As discussed earlier, part of this overestimation of the CLM3.5 simulation outputs may be due to the underrepresentation of soil moisture maxima from the Mesonet interpolated fields and errors associated with modeling. The CLM–CMORPH experiment does not perform as well as the experiments based on the other two remote sensing rainfall products, since the respective time series is characterized by significant bias with respect to the measured soil moisture. This is mainly due to an actual overestimation in summer-season precipitation by the CMORPH product (described in Anagnostou et al. 2010), which is consistent with results of previous studies (e.g., Tian et al. 2007). The 2006 summer period is generally drier than 2004, as indicated by the accumulated rainfall amounts and the soil moisture magnitudes shown in Fig. 3b. Here, CLM–MESONET, CLM–TRMM, and CLM–NEXRAD experiments tend to slightly underestimate soil moisture magnitudes most of the time (periods with no or almost no rainfall). A closer look at the soil moisture temporal evolution during the early days of June 2006 (Fig. 3b) reveals that a portion of this bias originates from the preceding period (e.g., the spring season, which is not depicted here). However, this propagated spring-induced bias is steadily increasing with time as the simulation advances to July and August. On the other hand, these CLM3.5 simulations seem to agree with the Mesonet measurements during the peak periods of heavy rainfall and increased soil moisture, although the latter should be mainly ascribed to the nonpreservation of high-frequency modulations in interpolated soil moisture that relate to short-term (convective) heavy rainfall events. CLM–CMORPH again exhibits significant positive bias with respect to the measured soil moisture.
Further, one could argue that the observed underestimation of the area-average near-surface soil moisture magnitude by most CLM3.5 simulations during summer 2006 could be attributed to the long-term and continuous nature of the simulations. If such an argument were true, then we would expect to observe a respective—lower in magnitude but still evident—bias in the summer 2005. However, this is not the case here, as the 2005 warm season is characterized by a good correspondence between measured and model-predicted soil moisture (except for the CLM–CMORPH experiment), very similar to what we observed in Fig. 3a for the 2004 warm season (graph not shown). It is worth mentioning that summer 2005 was also a relatively wet period for the Oklahoma region with a Mesonet summer rainfall accumulation of about 105 mm, a value very close to the 2004 summer period rainfall accumulation indicated in Fig. 3a.
A further difference between the two summer seasons pertains to the error magnitude. A cross examination of Figs. 4, 5 with respect to the area-average modeling- and rainfall-induced errors reveals that both types of error are lower in 2004 than 2006 (except for the case of the TRMM-forced RMSE, which is about 0.01 for both summer seasons). The area-average modeling RMSE in 2004 is approximately 0.015, as opposed to an almost double value (0.026) for 2006. The relatively low NEXRAD-forced (0.0055) RMSE and relatively high CMORPH-forced (0.04) RMSE in summer 2004 also appear significantly increased in summer 2006 (0.009 and 0.053, respectively). These differences in RMSEs could be attributed to the increased bias characterizing the model-predicted soil moisture during the warm season of 2006 (discussed earlier) and not be entirely associated with the drier conditions that prevailed over the same period.
Another noteworthy feature observed in Figs. 4, 5 is the relatively similar values of modeling- and rainfall-induced errors in soil moisture prediction. If the CLM–CMORPH experiment is excluded, which is characterized by large values of grid-cell RMSEs both in 2004 and 2006, then CLM–TRMM and CLM–NEXRAD rainfall-induced errors are of the same order with the modeling error (below 0.03). This is examined more in depth in Figs. 6, 7, where the rainfall-forced soil moisture RMSE is plotted against the model-induced soil moisture RMSE for all grid cells for the warm seasons of 2004 and 2006, respectively. These grid-cell distributions reveal that for most areas in Oklahoma, the 2004 CLM–TRMM rainfall-induced error is quite similar to the modeling error and that the respective CLM–NEXRAD error is slightly lower than the modeling error. In the 2006 summer season, the grid-cell CLM–TRMM and CLM–NEXRAD distributions show similar variability, with a tendency toward higher modeling errors. Clear differences emerge after averaging up to the domain scale. Both CLM–TRMM and CLM–NEXRAD average rainfall-induced errors are lower than the respective model-induced error in summer 2006, as opposed to similar or slightly different errors in summer 2004 (differences are shown to be statistically significant at the 99% confidence interval). CLM–CMORPH rainfall-induced error, however, is much higher than the modeling error in soil moisture prediction for most areas of the Oklahoma domain and in both seasons under study, which is yet associated with the prescribed CMORPH rainfall bias manifestation during the warm period of the year. It is noted here that the modeling error estimates are expected to be slightly overestimated, since they are subject to the source of uncertainty associated with the Mesonet soil moisture observation and spatial interpolation errors. However, our results and conclusions are not affected, since this uncertainty has a similar effect on the modeling error of both seasons under comparison in the current study.
The results discussed earlier are generally consistent with results presented in Fig. 8 in Hossain and Anagnostou (2005a) that were based on numerical experiments performed with ensemble-based techniques (see section 1). Direct quantitative comparisons between the results of the two studies cannot be performed, since different statistical approaches are used. Hossain and Anagnostou (2005a) addressed the issue of the relative contributions of modeling and rainfall uncertainties to the total uncertainty, which we have not directly tackled in the current study. However, both studies isolate and intercompare modeling and rainfall uncertainties. Specifically, Fig. 8 in Hossain and Anagnostou (2005a) indicates that the modeling and rainfall uncertainties are approximately of the same magnitude when the modeling accuracy is high, as opposed to the case of low modeling accuracy that is characterized by a rainfall error contribution much lower than the modeling error. Our results are equivalent, in the sense that the moist 2004 summer season represents our physically based case of higher modeling accuracy, which is characterized by almost similar values of domain-averaged modeling and rainfall contributed errors in the soil moisture simulations (TRMM and NEXRAD experiments; Figs. 6a,c, respectively), and the dry 2006 summer season corresponds to the case of lower modeling efficiency, demonstrating lower TRMM and NEXRAD domain-scale rainfall-induced errors in soil moisture simulations than the actual domain-scale modeling error (Figs. 7a,c, respectively).
c. Effect of scale on rainfall-to-soil moisture error propagation
An important addition to the issues addressed so far is the investigation of the effect of spatial scale on the rainfall error propagation in soil moisture prediction. The horizontal grid of 0.25° × 0.25° used in this study is the high-resolution grid currently utilized by the most recent spaceborne rainfall retrievals. However, past remotely sensed precipitation estimates as well as global model rainfall outputs use coarser spatial resolutions [see, e.g., the datasets used in Gottschalck et al. (2005); Ebert et al. (2007)]. Thus, it is important to examine the rainfall-induced error effect on the current LSM simulations of soil moisture rescaled to coarser horizontal grids. For that reason, the initial grid of 0.25° × 0.25° was postaggregated to larger-scale domains, for example, 0.5° × 0.5°, 1° × 1°, 1.25° × 2.75°, and 2.5° × 2.75°. All measured rain rates (from Mesonet, TRMM, CMORPH, and NEXRAD) as well as near-surface soil moisture (as measured from Mesonet and as predicted from the four CLM3.5 experiments) were interpolated to the aforementioned spatial scales. A point to note is that the methodology applied here imposes some limitations, in the sense that no actual simulations were performed at the coarser scales under consideration. The resolution effect on modeling is a major issue that merits detailed investigation that is beyond the scope of this study.
Figure 8 depicts the scale dependence of the relative rainfall forcing RMSE in soil moisture prediction as well as the ratio of this relative RMSE to the respective relative rain-rate RMSE, for each one of the three remote sensing products and respective model experiments and for both seasons under study. This ratio is an objective measure of the propagation of the actual error in rainfall estimates (model input error) to the soil moisture estimates (model output error) through the CLM3.5 simulations. The scale effect is evident in all cases. With respect to the rainfall-induced uncertainty in soil moisture alone (Fig. 8, left panels), we observe a decreasing trend with coarser spatial resolutions for all three experiments and for both seasons that could be attributed to the smoothing effect intrinsically associated with the aggregate nature of larger-scale fields. However, if this uncertainty is normalized by the respective error in rain rate (Fig. 8, right panels), then we clearly notice an increasing propagation error as the spatial resolution is getting lower for all three remotely sensed rainfall products and respective CLM3.5 experiments. The latter result further implies that the error in rainfall estimation decreases with coarser spatial resolutions and actually at a rate higher than the decreasing rate of the rainfall-induced uncertainty in soil moisture mentioned earlier. We expect that this gradient would be even stronger if land surface modeling were to be performed at the respective resolutions due to nonlinear hydrological processes. The fact that the rainfall propagation error grows with coarser spatial resolution exemplifies the need for using higher spatial resolution rainfall retrievals in the prediction of hydrological variables and underlines the importance of the GPM mission objectives.
Breaking the results of Fig. 8 down to each precipitation product and respective CLM3.5 experiment, we note high consistency with results presented in previous subsections. NEXRAD-related errors and error propagation ratios are the lowest observed for all scales under consideration (less than 10% and 4%, respectively, for both summer seasons), and the distribution for each scale is practically uniform around the median value. The respective TRMM-related error properties are similar with the exception of slightly higher values (less than 15% and 4%, respectively). CMORPH-related errors and error propagation ratios are much higher (at the order of 20%–30% and up to 8%, respectively), while the results are skewed toward the lower quartile values (especially in the case of summer 2004). Between the two different years, the summer season 2004 exhibits the lowest error magnitudes and error propagation ratios in accordance with the aforementioned positive effect of moisture abundance on rainfall-induced soil moisture uncertainty.
5. Summary and discussion
The current study presented an in-depth investigation of the properties of remotely sensed rainfall error propagation in the prediction of near-surface soil moisture from a land surface model. A plethora of data, including in situ measurements and remotely sensed retrievals, used either as forcing or reference for LSM simulations, facilitated a detailed analysis of the interaction between rainfall-induced uncertainty and modeling uncertainty and their effect on land surface state predictability. The study examined and quantified both sources of uncertainty in the simulation of soil moisture fields from an offline LSM (CLM3.5) forced by three different sources of remotely sensed precipitation estimates: two satellite sources (TRMM and CMORPH) and one from ground radar (NEXRAD). The assessment of both sources of error in soil moisture prediction was performed for the warm seasons of 2004 and 2006 and through the use of in situ measurements of soil moisture, rainfall, and other meteorological variables on a small domain in the midwestern United States capturing the state of Oklahoma (Oklahoma Mesonet). The study also presented a rigorous benchmarking of the Mesonet network as to its accuracy in deriving area rainfall estimates at the resolution of satellite products (0.25°, 3 h) based on comparisons against the most definitive measurements of a smaller-yet-denser network in southwestern Oklahoma (Micronet).
Our results indicated a dependence of the CLM3.5 efficiency in predicting near-surface (at 5-cm depth) soil moisture fields on the rainfall spatiotemporal climatology; model experiments showed better performance in the case of the relatively moist summer 2004, as opposed to a lowered performance skill during the relatively dry summer 2006. Further, moister areas within the Oklahoma region were associated with reduced rainfall-based error with respect to drier areas. All error magnitudes (e.g., modeling and remotely sensed rainfall-induced errors) were also shown to be lower in 2004 than in 2006. NEXRAD- and TRMM-induced errors in near-surface soil moisture were generally of low magnitude (at the order of 0.02–0.03) and comparable to each other, whereas the respective CMORPH-induced error was much higher overall (at the order of 0.05–0.06), due to excessive positive bias with respect to measured soil moisture (originating from a respective bias in rainfall estimation). An intercomparison between rainfall- and modeling-induced errors verified the results by Hossain and Anagnostou (2005a); both errors were of similar magnitude in the case of high modeling accuracy (e.g., warm season 2004), while rainfall-induced error was lower when the model’s efficiency skill was relatively low (e.g., warm-season 2006). Furthermore, a statistical evaluation of the scale effect on error properties revealed an increasing trend of the error propagation ratio (ratio of error in model-predicted soil moisture to the actual error in rainfall estimation) with coarser spatial resolutions.
Although the study presented useful indications about the error propagation properties of rainfall in the simulation of soil moisture, it is limited in terms of the rainfall products investigated and LSM models. There are several satellite techniques providing high-resolution global-scale products with varying error characteristics and resolutions that need to be investigated in terms of their efficiency in the prediction of soil moisture variability. Furthermore, a point to note is that most satellite retrievals, including the two techniques presented herein, are undergoing changes aimed at improving their rain estimation error characteristics (e.g., the CMORPH technique is currently undergoing several modifications in the way morphing is performed to account for the summertime positive bias). The model used here to facilitate our investigation is only one of the several land surface schemes currently included in major land data assimilation systems. The models exhibit differences in the parameterizations devised to represent the land–atmosphere interaction processes, which would subsequently affect in a nonlinear way the propagation of precipitation error in the prediction of hydrologic parameters (including the soil moisture studied herein). Those aspects are subject to future research studies.
Acknowledgments
This work was supported by EU Marie Curie Excellence Grant project PreWEC MEXT-CT-2006-038331 and by NASA Precipitation Measurement Mission Award NNX07AE31G. E. I. Nikolopoulos was supported by a NASA Earth System Science Graduate Fellowship.
REFERENCES
Anagnostou, E. N., Maggioni V. , Nikolopoulos E. I. , Meskele T. , Hossain F. , and Papadopoulos A. , 2010: Benchmarking high-resolution global satellite rainfall products to radar and rain-gauge rainfall estimates. IEEE Trans. Geosci. Remote Sens., 48 , 1667–1683. doi:10.1109/TGRS.2009.2034736.
Beljaars, A. C. M., Viterbo P. , Miller M. J. , and Betts A. K. , 1996: The anomalous rainfall over the United States during July 1993: Sensitivity to land surface parameterization and soil moisture anomalies. Mon. Wea. Rev., 124 , 362–383.
Betts, A. K., 2004: Understanding hydrometeorology using global models. Bull. Amer. Meteor. Soc., 85 , 1673–1688.
Beven, K. J., and Kirkby M. J. , 1979: A physically based, variable contributing area model of basin hydrology. Hydrol. Sci. Bull., 24 , 43–69.
Brock, F. V., Crawford K. C. , Elliott R. L. , Cuperus G. W. , Stadler S. J. , Johnson H. L. , and Eilts M. D. , 1995: The Oklahoma Mesonet: A technical overview. J. Atmos. Oceanic Technol., 12 , 5–19.
Cosgrove, B. A., and Coauthors, 2003: Land surface model spin-up behaviour in the North American Land Data Assimilation System (NLDAS). J. Geophys. Res., 108 , D228845. doi:10.1029/2002JD003316.
Drusch, M., and Viterbo P. , 2007: Assimilation of screen-level variables in ECMWF’s Integrated Forecast System: A study on the impact on the forecast quality and analyzed soil moisture. Mon. Wea. Rev., 135 , 300–314.
Ebert, E. E., Janowiak J. E. , and Kidd C. , 2007: Comparison of near-real-time precipitation estimates from satellite observations and numerical models. Bull. Amer. Meteor. Soc., 88 , 47–64.
Entekhabi, D., and Coauthors, 2004: The Hydrosphere State (Hydros) Satellite Mission: An earth system pathfinder for global mapping of soil moisture and land freeze/thaw. IEEE Trans. Geosci. Remote Sens., 42 , 2184–2195.
Fischer, E. M., Seneviratne S. I. , Vidale P. L. , Lüthi D. , and Schär C. , 2007: Soil moisture–atmosphere interactions during the 2003 European summer heat wave. J. Climate, 20 , 5081–5099.
Fulton, R. A., Breidenbach J. P. , Seo D. J. , Miller D. A. , and O’Bannon T. , 1998: The WSR-88D rainfall algorithm. Wea. Forecasting, 13 , 377–395.
Gottschalck, J., Meng J. , Rodell M. , and Houser P. , 2005: Analysis of multiple precipitation products and preliminary assessment of their impact on Global Land Data Assimilation System land surface states. J. Hydrometeor., 6 , 573–598.
Hossain, F., and Anagnostou E. N. , 2004: Assessment of current passive-microwave- and infrared-based satellite rainfall remote sensing for flood prediction. J. Geophys. Res., 109 , D07102. doi:10.1029/2003JD003986.
Hossain, F., and Anagnostou E. N. , 2005a: Numerical investigation of the impact of uncertainties in satellite rainfall estimation and land surface model parameters on simulation of soil moisture. Adv. Water Resour., 28 , 1336–1350.
Hossain, F., and Anagnostou E. N. , 2005b: Using a multi-dimensional satellite rainfall error model to characterize uncertainty in soil moisture fields simulated by an offline land surface model. Geophys. Res. Lett., 32 , L15402. doi:10.1029/2005GL023122.
Hossain, F., and Anagnostou E. N. , 2006: A two-dimensional satellite rainfall error model. IEEE Trans. Geosci. Remote Sens., 44 , 1511–1522.
Hossain, F., and Huffman G. J. , 2008: Investigating error metrics for satellite rainfall data at hydrologically relevant scales. J. Hydrometeor., 9 , 563–575.
Huffman, G. J., and Coauthors, 2007: The TRMM Multisatellite Precipitation Analysis (TMPA): Quasi-global, multiyear, combined-sensor precipitation estimates at fine scales. J. Hydrometeor., 8 , 38–55.
Joyce, R. J., Janowiak J. E. , Arkin P. A. , and Xie P. , 2004: CMORPH: A method that produces global precipitation estimates from passive microwave and infrared data at high spatial and temporal resolution. J. Hydrometeor., 5 , 487–503.
Kerr, Y. H., Waldteufel P. , Wigneron J. P. , Martinuzzi J. M. , Font J. , and Berger M. , 2001: Soil moisture retrieval from space: The Soil Moisture and Ocean Salinity (SMOS) mission. IEEE Trans. Geosci. Remote Sens., 39 , 1729–1735.
Koster, R. D., and Suarez M. J. , 2001: Soil moisture memory in climate models. J. Hydrometeor., 6 , 558–570.
Koster, R. D., and Coauthors, 2004a: Realistic initialization of land surface states: Impacts on subseasonal forecast skill. J. Hydrometeor., 5 , 1049–1063.
Koster, R. D., and Coauthors, 2004b: Regions of strong coupling between soil moisture and precipitation. Science, 305 , 1138–1140.
Kumar, S. V., and Coauthors, 2006: Land information system: An interoperable framework for high resolution land surface modeling. Environ. Modell. Software, 21 , 1402–1415.
Lawrence, P. J., and Chase T. N. , 2007: Representing a new MODIS consistent land surface in the Community Land Model (CLM 3.0). J. Geophys. Res., 112 , G01023. doi:10.1029/2006JG000168.
Leese, J., Jackson T. , Pitman A. , and Dirmeyer P. , 2001: Meeting summary: GEWEX/BAHC International Workshop on Soil Moisture Monitoring, Analysis, and Prediction for Hydrometeorological and Hydroclimatological Applications. Bull. Amer. Meteor. Soc., 82 , 1423–1430.
Lin, Y., and Mitchell K. E. , 2005: The NCEP stage II/IV hourly precipitation analyses: Development and applications. Preprints, 19th Conf. on Hydrology, San Diego, CA, Amer. Meteor. Soc., 1.2. [Available online at http://ams.confex.com/ams/Annual2005/techprogram/paper_83847.htm].
Mitchell, K. E., and Coauthors, 2004: The multi-institution North American Land Data Assimilation System (NLDAS): Utilizing multiple GCIP products and partners in a continental distributed hydrological modeling system. J. Geophys. Res., 109 , D07S90. doi:10.1029/2003JD003823.
Niu, G-Y., Yang Z-L. , Dickinson R. E. , and Gulden L. E. , 2005: A simple TOPMODEL-based runoff parameterization (SIMTOP) for use in global climate models. J. Geophys. Res., 110 , D21106. doi:10.1029/2005JD006111.
Oleson, K. W., and Coauthors, 2004: Technical description of the Community Land Model (CLM). NCAR Tech. Note NCAR/TN-461+STR, 173 pp.
Oleson, K. W., and Coauthors, 2008: Improvements to the Community Land Model and their impact on the hydrological cycle. J. Geophys. Res., 113 , G01021. doi:10.1029/2007JG000563.
Papadopoulos, A., Serpetzoglou E. , and Anagnostou E. N. , 2008: Improving NWP through radar rainfall-driven land surface parameters: A case study on convective precipitation forecasting. Adv. Water Res., 31 , 1456–1469.
Reichle, R. H., Koster R. D. , Dong J. , and Berg A. A. , 2004: Global soil moisture from satellite observations, land surface models, and ground data: Implications for data assimilation. J. Hydrometeor., 5 , 430–442.
Reichle, R. H., Koster R. D. , Liu P. , Mahanama S. P. P. , Njoku E. G. , and Owe M. , 2007: Comparison and assimilation of global soil moisture retrievals from the Advanced Microwave Scanning Radiometer for the Earth Observing System (AMSR-E) and the Scanning Multichannel Microwave Radiometer (SMMR). J. Geophys. Res., 112 , D09108. doi:10.1029/2006JD008033.
Robock, A., Vinnikov K. Y. , Srinivasan G. , Entin J. K. , Hollinger S. E. , Speranskaya N. A. , Liu S. , and Namkhai A. , 2000: The Global Soil Moisture Data Bank. Bull. Amer. Meteor. Soc., 81 , 1281–1299.
Robock, A., and Coauthors, 2003: Evaluation of the North American Land Data Assimilation System over the Southern Great Plains during the warm season. J. Geophys. Res., 108 , 8846. doi:10.1029/2002JD003245.
Rodell, M., and Coauthors, 2004: The global land data assimilation system. Bull. Amer. Meteor. Soc., 85 , 381–394.
Rodell, M., Houser P. R. , Berg A. A. , and Famiglietti J. S. , 2005: Evaluation of 10 methods for initializing a land surface model. J. Hydrometeor., 6 , 146–155.
Schaake, J. C., and Coauthors, 2004: An intercomparison of soil moisture fields in the North American Land Data Assimilation System (NLDAS). J. Geophys. Res., 109 , D01S90. doi:10.1029/2002JD003309.
Smith, E. A., and Coauthors, 2007: International Global Precipitation Measurement (GPM) program and mission: An overview. Measuring Precipitation from Space: EURAINSAT and the Future, V. Levizzani, P. Bauer, and F. J. Turk, Eds., Springer, 611–654.
Sorooshian, S., Hsu K-L. , Gao X. , Gupta H. V. , Imam B. , and Braithwaite D. , 2000: Evaluation of PERSIANN system satellite–based estimates of tropical rainfall. Bull. Amer. Meteor. Soc., 81 , 2035–2046.
Stöckli, R., and Coauthors, 2008: Use of FLUXNET in the Community Land Model development. J. Geophys. Res., 113 , G01025. doi:10.1029/2007JG000562.
Tian, X., Xie Z. , and Dai A. , 2008: A land surface soil moisture data assimilation system based on the dual-UKF method and the Community Land Model. J. Geophys. Res., 113 , D14127. doi:10.1029/2007JD009650.
Tian, Y., Peters-Lidard C. D. , Choudhury B. J. , and Garcia M. , 2007: Multitemporal analysis of TRMM-based satellite precipitation products for land data assimilation applications. J. Hydrometeor., 8 , 1165–1183.
Walker, J. P., and Houser P. , 2004: Requirements of a global near-surface soil moisture satellite mission: Accuracy, repeat time, and spatial resolution. Adv. Water Resour., 27 , 785–801.
Walker, J. P., Houser P. , and Reichle R. , 2003: New technologies require advances in hydrologic data assimilation. Eos, Trans. Amer. Geophys. Union, 84 .doi:10.1029/2003EO490002.
Schematic of the Oklahoma Mesonet and Micronet networks. The locations of the Mesonet stations are indicated by the black dots. The centers of the cells of the 0.25° grid shown in the same graph represent the locations where soil moisture data are extracted based on the application of gauge interpolation technique. The ARS Little Washita River watershed is also expanded in a bigger diagram showing the locations of the Micronet stations (picture obtained online at http://ars.mesonet.org/sites/). The stations in red have been retired since 2005.
Citation: Journal of Hydrometeorology 11, 3; 10.1175/2009JHM1166.1
Comparison between high-resolution (Micronet) and low-resolution (Mesonet) rainfall forcing for the 2004 warm season. (a),(top) Time series of average Mesonet and Micronet rain accumulation (mm) and Mesonet rain rate (mm h−1) for the respective Mesonet–Micronet grid cell. (bottom) Time series of observed near-surface soil moisture (Mesonet) and predicted near-surface soil moisture values (CLM–MESONET and CLM–MICRONET experiments) for the Mesonet–Micronet grid cell. (b) Absolute relative errors in near-surface soil moisture for CLM–MICRONET vs the respective error for CLM–MESONET, for each 3-h record of the Mesonet–Micronet grid cell.
Citation: Journal of Hydrometeorology 11, 3; 10.1175/2009JHM1166.1
Time series of observed rain rate (mm h−1), rain accumulation (mm), and soil moisture (at 5 cm) from the Mesonet network and predicted soil moisture from the four CLM3.5 experiments, for the summer periods of (a) 2004 and (b) 2006. In all cases, the average value of the gridded domain shown in Fig. 1 is depicted.
Citation: Journal of Hydrometeorology 11, 3; 10.1175/2009JHM1166.1
Near-surface soil moisture grid-cell RMSEs vs ΔRi for (a) CLM–MESONET, (b) CLM–TRMM, (c) CLM–CMORPH, and (d) CLM–NEXRAD, for the summer period 2004. The dashed lines indicate the RMSE of the respective mean values (averages for the entire gridded domain), which is independent of ΔRi.
Citation: Journal of Hydrometeorology 11, 3; 10.1175/2009JHM1166.1
As in Fig. 4, but for summer 2006.
Citation: Journal of Hydrometeorology 11, 3; 10.1175/2009JHM1166.1
Near-surface soil moisture RMSE of (a) TRMM-, (b) CMORPH-, and (c) NEXRAD-forced experiments (e.g., rainfall-induced error) vs near-surface soil moisture RMSE of Mesonet-forced experiment (e.g., model-induced error) for each grid cell, for the 2004 summer period. The red marker denotes the relationship between the RMSEs of the respective mean values (averages for the entire gridded domain). The length of each red bar represents the respective 99% confidence interval.
Citation: Journal of Hydrometeorology 11, 3; 10.1175/2009JHM1166.1
As in Fig. 6, but for summer 2006.
Citation: Journal of Hydrometeorology 11, 3; 10.1175/2009JHM1166.1
Box-and-whisker plots indicating the scale effect on rainfall error propagation in soil moisture prediction for the warm seasons of 2004 (blue colors) and 2006 (red colors): (left) the relative rainfall-induced RMSE in near-surface soil moisture and (right) the ratio of this relative RMSE (rRMSE) over the relative RMSE in rain rate, for (top) CLM–TRMM, (middle) CLM–CMORPH, and (bottom) CLM–NEXRAD for each of the prescribed scales. The boxes have lines at the lower and upper quartile values and the median value, and the whiskers are the dashed lines extending from each end of the boxes showing the extent of the rest of the data.
Citation: Journal of Hydrometeorology 11, 3; 10.1175/2009JHM1166.1
High-resolution surface datasets used in CLM3.5.
