Assimilation experiments have been performed with the Weather Research and Forecasting (WRF) model’s three-dimensional variational data assimilation (3DVAR) scheme to assess the impacts of NASA’s Quick Scatterometer (QuikSCAT) near-surface winds, and Special Sensor Microwave Imager (SSM/I) wind speed and total precipitable water (TPW) on the analysis and on short-range forecasts over the Indian region. The control (without satellite data) as well as WRF 3DVAR sensitivity runs (which assimilated satellite data) were made for 48 h starting daily at 0000 UTC during July 2006. The impacts of assimilating the different satellite dataset were measured in comparison to the control run, which does not assimilate any satellite data. The spatial distribution of the forecast impacts (FIs) for wind, temperature, and humidity from 1-month assimilation experiments for July 2006 demonstrated that on an average, for 24- and 48-h forecasts, the satellite data provided useful information. Among the experiments, WRF wind speed prediction was improved by QuikSCAT surface wind and SSM/I TPW assimilation, while temperature and humidity prediction was improved due to the assimilation of SSM/I TPW. The rainfall prediction has also been improved significantly due to the assimilation of SSM/I TPW, with the largest improvement seen over the west coast of India. Through an improvement of the surface wind field, the QuikSCAT data also yielded a positive impact on the precipitation, particularly for day 1 forecasts. In contrast, the assimilation of SSM/I wind speed degraded the humidity and rainfall predictions.
Socioeconomic aspects of life in India are highly dependent on both the intensity and distribution of summer monsoon rainfall. Therefore, providing accurate weather forecasts using numerical weather prediction (NWP) models during the monsoon season is of primary importance within the scientific community. In recent years, most of the meteorological agencies and researchers have depended on guidance from NWP models in issuing rainfall forecasts 1–2 days in advance. The simulation of the Indian summer monsoon (ISM) has been the focus of many modeling studies due to its anomalous characteristics in the tropical circulation (Hahn and Manabe 1975; Fennessy et al. 1994; Ashok et al. 1998; Chandrasekhar et al. 1999; Eitzen and Randall 1999; Das et al. 2002; Ratnam and Kumar 2005). The numerical weather forecasts exhibit uncertainties, which can be due to errors in the initial conditions, the representation of physical processes, or the computational precision used in the model. Even though progress has been made in terms of computational speed, observation networks, NWP techniques, and physical parameterizations, weather forecasts on the regional scale have not yet reached the required accuracy (Kalnay 2003).
Understanding the errors in NWP models can only be achieved by extensive verification of these models for various synoptic conditions and by conducting impact studies using better initial conditions (through data assimilation). Rakesh et al. (2007, 2009, manuscript submitted to Meteor. Appl.) evaluated the precipitation skill of the widely used fifth-generation Pennsylvania State University–National Center for Atmospheric Research (PSU–NCAR) Mesoscale Model (MM5) over the Indian region and found that the skill of the precipitation forecasts is still not satisfactory. They also suggested the need for improvements in precipitation forecasts through the choice of better physics options by sensitivity studies and more accurate initial conditions by the assimilation of observations.
A continuing difficulty with respect to the improvements in forecasts at smaller spatial scales by mesoscale models relates to the fact that observational information is limited and inaccurate, especially in data-sparse areas such as large oceans and deserts. Data assimilation has been recognized as a useful way to obtain better “consistent” initial conditions for NWP (Kalnay 2003). Recent improvements in remote sensing technology make it possible to observe the atmosphere in areas where conventional observations are sparse. A number of case studies (Zou and Xiao 2000; Pu et al. 2002; Harasti et al. 2004; Chen 2007; Zhang et al. 2007; Singh et al. 2008a) have shown that remote sensing of data over the oceans can improve tropical cyclone initialization and prediction. Similarly, other studies (Fan and Tilley 2005; Chou et al. 2006; Powers 2007; Singh et al. 2008b; Rakesh et al. 2009) have shown that satellite data assimilation improved the forecasted meteorological features associated with various weather systems. A major drawback of such case studies is the limited number of forecast samples and the statistics resulting from them may not be robust enough to reach a firm conclusion. Zapotocny et al. (2007) studied the impacts of various satellite and in situ data in the National Centers for Environmental Prediction (NCEP) Global Data Assimilation System (GDAS) for two different seasons. Their results showed a positive impact from both conventional in situ and remotely sensed satellite data during both seasons in both hemispheres.
The present study used the maturing Weather Research and Forecasting (WRF; Skamarock et al. 2005) model, which is the successor to MM5, to investigate the potential impacts of the National Aeronautics and Space Administration’s (NASA’s) Quick Scatterometer (QuikSCAT) surface winds, Special Sensor Microwave Imager (SSM/I) wind speed, and total precipitable water (TPW) on short-range forecasts. The WRF model was used due to its vastly growing popularity within the mesoscale modeling community and the fact that it is in the developmental stage. The developmental applications of WRF have primarily been in the midlatitudes, and to date the studies over tropical regions are comparatively less in number. In light of this, we used the WRF for short-range forecast applications during the 2006 monsoon over the Indian region for the satellite data impact study. The benefits to the modeling community of the present study, as compared to case studies, are due to the large numbers of forecasts generated during this experiment, which make statistical evaluation an appropriate tool for identifying model discrepancies. These results provide helpful information for further developments in numerical models. One of the objectives of the present study is to explore the short-range forecast skill of the WRF model over the Indian region. The work described in this paper is also relevant in quantifying the potential impacts of QuikSCAT and SSM/I observations on the WRF short-range forecasts. The paper is structured as follows: section 2 describes the satellite data assimilated in this study; descriptions of the WRF model, the assimilation methodology, and the design of the numerical experiments are given in section 3; the initialization and simulation results of the study are shown in section 4; section 5 discusses the sensitivity of assimilation results to the cumulus parameterization; and the paper is summarized in section 6.
2. Data used for assimilation
a. QuikSCAT surface winds
Launched in June 1999, the QuikSCAT orbits the earth at an altitude of 800 km once every 101 min (Shirtliffe 1999). Having a swath width of approximately 1800 km, the SeaWinds instrument aboard the QuikSCAT satellite operates at the 13.4-GHz Ku band. The accuracy of the measured ocean surface wind reaches 2 m s−1 in speed and 20° in direction for winds of 3–20 m s−1 and 10% for winds of 20–30 m s−1 (Shirtliffe 1999). Information from independent data sources (e.g., numerical models) is needed to remove the ambiguity in the direction determination. While QuikSCAT observations can be contaminated by a rainy atmosphere (Weissman et al. 2002; Sharp et al. 2002; Pasch et al. 2003; Leidner et al. 2003; Hoffman and Leidner 2005), recent assimilation studies (Atlas et al. 2001; Leidner et al. 2003; Goerss and Hogan 2006; Zhang et al. 2007; Chen 2007; Zapotocny et al. 2007; Singh et al. 2008a) have shown that QuikSCAT data have a positive impact on the analysis and prediction of weather systems. The utility of QuikSCAT winds in extratropical cyclone forecasting and marine weather prediction at the National Oceanographic and Atmospheric Administration’s (NOAA) Ocean Prediction Center (OPC) is documented by Von Ahn et al. (2006), while Chelton et al. (2006) and Atlas et al. (2001) described the utility of scatterometer winds in general marine weather forecasting applications. The use of QuikSCAT data in tropical cyclone analysis and forecasting at the National Hurricane Center (NHC) is described by Brennan et al. (2009).
b. SSM/I wind and TPW
The SSM/I (Hollinger 1989) is a conical scanning, four-frequency, linearly polarized, seven-channel passive microwave radiometer, which has operated on board Defense Meteorological Satellite Program (DMSP) satellites since June 1987. This polar-orbiting satellite has a period of approximately 102 min, a near-constant incidence angle of 53°, a mean altitude of approximately 830 km, and a swath width of about 1400 km. Like QuikSCAT, SSM/I data are available under both clear and cloudy conditions but can be contaminated by precipitation. The retrieved TPW and sea surface winds from the DMSP F13 satellite are used in this study. The TPW and the sea surface wind were both derived from brightness temperatures using Wentz’s algorithm (Wentz 1997) in rain-free areas. The resolution of SSM/I data is 25 km, which is same as that of QuikSCAT winds. The SSM/I provides only wind speed, while QuikSCAT provides both wind speed and direction. The positive impacts of SSM/I data on the NWP analysis and prediction have been shown by Gerard and Saunders (1999), Xiao et al. (2000), Chen et al. (2004), and Kelly et al. (2008).
3. WRF model and assimilation methodology
a. WRF model
The forecast model used is the Weather Research and Forecasting (Skamarock et al. 2005) model, version 2.2. WRF is a next-generation mesoscale NWP system designed to serve both operational forecasting and atmospheric research needs. It is a limited-area, nonhydrostatic, primitive-equation model with multiple options for various physical parameterization schemes. There are two dynamics solvers within the WRF software framework: the Advanced Research WRF (ARW) solver, developed primarily at NCAR, and the Nonhydrostatic Mesoscale Model (NMM) solver, developed at NCEP. We have used the ARW dynamic solver for the present study. This version employs Arakawa C-grid staggering for the horizontal grid and a fully compressible system of equations. The terrain-following hydrostatic pressure coordinate with vertical grid stretching was followed in the vertical. The time-split integration uses a third-order Runge–Kutta scheme with a smaller time step for acoustic and gravity wave modes. The WRF physical options used in this study consisted of the WRF single-moment six-class (WSM6) graupel scheme for microphysics, which is similar to that used by Lin et al. (1983); the new Kain–Fritsch (KF; Kain 2004) cumulus convection parameterization scheme; and the Yonsei University (YSU) planetary boundary layer scheme (Hong and Dudhia 2003). The Rapid Radiative Transfer Model (RRTM; Mlawer et al. 1997) and the Dudhia scheme (Dudhia 1989) were used for longwave and shortwave radiation, respectively. All experiments were conducted with a single domain (Fig. 1) consisting of 170 × 170 grid points with 30-km horizontal grid resolution. The model had 28 vertical levels with the top of the model atmosphere located at 50 hPa.
b. Assimilation methodology
The WRF three-dimensional variational data assimilation (3DVAR) system (Skamarock et al. 2005) was used in this study. The WRF 3DVAR evolved from the MM5 3DVAR system (Barker et al. 2004), but the basic software interface and coordinate framework were fully updated for the WRF model. The background covariances matrix was estimated using the so-called National Meteorological Center (NMC, which is now known as NCEP) method (Parrish and Derber 1992; Wu et al. 2002). The observation errors were assumed to be uncorrelated in space and time. Since observation errors were assumed to be uncorrelated, the observational error covariances matrices were simple diagonal, with QuikSCAT or SSM/I observation error variances as elements. In this study, these variances were taken as constant in space and time. The standard deviations for the QuikSCAT wind speed and direction were 1.4 m s−1 and 20°, respectively, while the standard deviations in SSM/I winds and TPW were 2.5 m s−1 and 0.2 g cm−2, respectively. Prior to data assimilation, all satellite data underwent quality-checking processes in order to reduce the possibility of assimilating bad observations. First, rain-contaminated data were excluded from the QuikSCAT and SSM/I data. The rainfall probability parameter (p) along with the QuikSCAT wind product are used to exclude the rain-contaminated observations from the QuikSCAT winds. The SSM/I retrievals were already flagged for the rain-contaminated pixels by the SSM/I data products generation team. Second, a gross error quality control was performed in which observations (from QuikSCAT and SSM/I) that differed from the model first guess by more than 5 times the observational errors were removed.
c. Design of numerical experiments
Five analyses (Table 1) were produced daily from 1 to 31 July 2006 at 0000 UTC using the WRF 3DVAR. The control 3DVAR experiments (CNT; Table 1), which only assimilated conventional radiosonde observations, were used for comparison purposes, as they did not assimilate any satellite data. The conventional radiosonde data were also assimilated in satellite data sensitivity experiments (QW, SW, SWV, and SWWV; details are given in Table 1) in order to ensure the dynamical consistency among all the analyses. Instead of directly using the NCEP final (FNL) analysis for the first guess (FG), 6-h WRF forecasts initialized using the NCEP FNL analysis were used as the FG for all 3DVAR experiments. The NCEP FNL analysis with 1° × 1° resolution was used for the model boundary conditions for all the experiments. Forty-eight-hour forecasts were made daily from 1 to 31 July 2006 at 0000 UTC with five (CNT, QW, SW, SWV, and SWWV) different sets of initial conditions.
4. Assimilation results
a. Overview of the fit to observations
The first comparison that we made can be described as a sanity check; the mean and root-mean-square errors (RMSEs) of the observed minus analysis (analysis departures), as well as the observed minus first guess (first-guess departures), were analyzed for the different experiments. In a successful assimilation, the analysis departures (O − A) are smaller than the first-guess departures (O − B); hence, the analysis better matches the observations. This is illustrated in Fig. 2, showing histograms of first-guess and analysis departures for QuikSCAT (Figs. 2a and 2b) and SWV (Figs. 2a and 2d) experiments. For QuikSCAT, the first-guess departures (O − B; Fig. 2a) have an RMSE of about 2.2 m s−1, while the analysis departures (O − A; Fig. 2b) have an RMSE of about 0.5 m s−1. The mean difference is reduced from −0.31 m s−1 (O − B) to 0.05 m s−1 (O − A). In the SSM/I TPW assimilation (SWV) case, the mean difference is reduced from 0.245 g cm−2 (O − B; Fig. 2c) to −0.002 g cm−2 (O − A; Fig. 2d), while RMSE is reduced from 0.387 g cm−2 (O − B; Fig. 2c) to 0.028 g cm−2 (O − A; Fig. 2d). For SW (not shown), the RMSE of the analysis departures is of the order of 1.5 m s−1 as opposed to 3.2 m s−1 for the first-guess departures. Overall, Fig. 2 confirms that 3DVAR is successful in bringing the analysis closer to the observations than the background.
b. Impacts of QuikSCAT and SSM/I data on the analysis
To see the sensitivity of different satellite data in the 3DVAR initial analysis, we computed the spatial distribution of root-mean-square sensitivity (RS; see the appendix) in wind speed (m s−1) and relative humidity (%) between the experimental (with satellite data) and control (CNT; without satellite data) cases. Figures 3a–c show the RS in initial wind speed between the QuikSCAT wind assimilation (QW) and CNT (Fig. 3a), SSM/I wind assimilation (SW) and CNT (Fig. 3b), and SSM/I TPW assimilation (SWV) and CNT (Fig. 3c). These results are obtained by comparing 30 samples of experimental analyses with the corresponding control analyses. The change in 950-hPa wind speed due to the assimilation of QuikSCAT (QW) and SSM/I (SW) data is mainly observed over the equatorial Indian Ocean and northern part of the Bay of Bengal (Figs. 3a and 3b). SW showed higher (of the order of 1 m s−1) sensitivity in the case of 950-hPa wind speed as compared to QW. Assimilation of SSM/I TPW (SWV) caused a 4%–6% change in the 950-hPa relative humidity and large sensitivity is seen over the northern part of the Arabian Sea, the Bay of Bengal, and the southwest equatorial Indian Ocean (Fig. 3c). To complement the results obtained in this section for the impacts of satellite data on initial analysis, we have carried out a significance test [Student’s t test (Dewberry 2004)] between the analysis from CNT and different assimilation experiments (Figs. 3d–f). It will give an idea about how significant the differences are in the analysis between CNT and the different assimilation experiments. The higher the value of t is, the greater is the confidence that there is a difference between the CNT and EXP analyses. Our analysis for the 950-hPa wind speed showed that in the CNT and QW analyses, significant differences are observed over the southwest Arabian Sea and scattered areas near the Indian coast (areas shaded in black show a difference that corresponds to more than 90% confidence level; see Fig. 3d). In the CNT and SW analyses, a large area of significant difference (above 90% confidence level) is observed over the equatorial Indian Ocean (Fig. 3e). Analysis of the 950-hPa relative humidity from CNT and SWV showed that a significant difference in humidity is observed over the Bay of Bengal and the Arabian Sea near the Indian coast (Fig. 3f).
c. Impacts of QuikSCAT and SSM/I data on the forecast
The bias or systematic errors are vital in diagnosing model discrepancies. In the forecasted fields, RMSE is considered to be a standard measure for evaluating model performance (Wang and Yongfu 2001). We analyzed the spatial distribution of the bias and RMSE in the 24- and 48-h predicted wind speed, temperature, and relative humidity. To see the impacts of different satellite data in WRF simulations, we computed the forecast impact (FI; see the appendix) based on the RMSEs of different assimilation experiments. A positive value of FI indicates improvement due to satellite data assimilation as compared to CNT, while a negative value of FI indicates degradation. The use of a normalized forecast impact parameter is advantageous for comparing the impacts of satellite data on different model-predicted variables irrespective of their original magnitudes. Hence, this parameter is quite useful in quantitatively discriminating the prediction of which variable at a particular level was most improved by the assimilation of a particular type of satellite data.
1) Wind speed
Spatial distributions of the biases and RMSEs in 24- and 48-h wind speed forecasts valid at 0000 UTC are computed by comparing 30 samples of CNT forecasts with corresponding QuikSCAT observations. We used the QuikSCAT observations falling in a time window of ±1 h near 0000 UTC for the comparison of model output. As compared with the observations, the 24-h WRF CNT forecasts underestimated the wind speed by 1.5–2 m s−1 over the southern Arabian Sea near peninsular India and this weakening extends slightly eastward in the 48-h forecasts (not shown). RMSEs of the order of 1–2 m s−1 are observed in the 24-h predicted wind speed from WRF CNT simulations over the oceanic region with larger values seen over the equatorial Indian Ocean south of peninsular India (Fig. 4a). The spatial distribution of FI from different assimilation experiments (Figs. 4b–e) shows that the assimilation of satellite data improved the predicted wind speed (more positive area as compared to negative area in Figs. 4b–e). Among the assimilation experiments, more positive area (representing improved prediction of wind speed) is seen in experiments in which QuikSCAT wind (QW) or SSM/I TPW (SWV) results are assimilated. The RMSEs in the 24-h forecasted wind speed by the WRF CNT simulation are increased by a magnitude of the order of 0.5 m s−1 in the 48-h forecasts (not shown). Similar to the 24-h forecast, experiments with the assimilation of satellite data improved the 48-h wind speed prediction, and the largest improvement is seen over the Bay of Bengal due to SSM/I TPW assimilation (SWV; not shown). The assimilation of TPW led to significant impacts on the winds. This change in the wind field is likely due to the indirect forcing of winds by model physics. For example, by adding moisture to the lower troposphere or removing moisture from the upper troposphere, the TPW can lead to wind adjustments.
To complement the results obtained from the verification of the model-predicted wind speed with QuikSCAT observations and considering that QuikSCAT observations are available only over the ocean, we compared the model-predicted wind field at 850 hPa with the NCEP-analyzed field. Comparison with the NCEP analysis showed very high RMSEs (more than 4 m s−1) in the WRF 24-h predicted wind speed near the southwest coast of India and the Bay of Bengal near Sri Lanka (Fig. 5a). Among the experiments, the one in which only the SSM/I TPW was assimilated (SWV), improved the 24-h predicted wind speed at 850 hPa (Figs. 5b–e). It is seen that the assimilation of SSM/I wind speeds resulted in a large degradation of wind speed prediction over the equatorial Indian Ocean and northeast India (Figs. 5c and 5e). Assimilation of SSMI TPW (SWV) showed large improvements (not shown) in the 48-h wind speed predictions over the equatorial Indian Ocean.
With the aim of verifying the model-predicted wind speed with an independent observation (which is not assimilated in this study), we compared the predicted wind speed at 10-m height with buoy/ship-observed wind speeds available from the International Comprehensive Ocean–Atmosphere Dataset (ICOADS; Woodruff et al. 1998; Worley et al. 2005). The conventional buoy/ship observations at 2-m height are converted to 10-m heights using the logarithmic wind profile relation followed by Mears et al. (2001) before using them to verify the model-predicted wind speed. It is observed from the scatterplot of 24-h predicted wind speeds by different experiments and buoy/ship observed wind speeds (Fig. 6) that the WRF forecasts showed an average RMSE of 3.5 m s−1 and a positive bias of 0.65 m s−1 over the oceanic region. The RMSE in forecasted wind speed slightly increased in the case of the 48-h forecast, while a significant change is not observed in the bias (not shown). Temporal variations in the forecasted wind speeds are analyzed by verifying them with buoy/ship observations over two selected locations: one over the Bay of Bengal (11.5°N, 81.5°E) and the other over the Arabian Sea (10.6°N, 72.4°E). It is clear that the RMSEs in the forecasted wind speeds from the WRF are larger over the Bay of Bengal (Fig. 7a) as compared to those over the Arabian Sea (Fig. 7b). The RMSEs in the predicted wind speeds are slightly less in the satellite data assimilation experiments as compared to the CNT at some forecast intervals over both areas (Fig. 7). The experiments in which SSM/I TPW are assimilated (SWV and SWWV) showed the lowest RMSEs in the predicted wind speeds for most of the forecast intervals over the Bay of Bengal while the experiments showed mixed results over the Arabian Sea (Fig. 7).
The model-predicted temperature is verified with the independent (not assimilated) observations from the Atmospheric Infrared Sounder (AIRS; Aumann et al. 2003). When the AIRS-retrieved thermodynamic profiles are compared to radiosondes (Fetzer et al. 2003; Tobin et al. 2006; Divakarla et al. 2006), RMSEs of 1 K in 1-km layers for temperature and 10%–15% in 2-km layers for relative humidity are found. The AIRS data used in this study are level 2 version 4.0 atmospheric temperature and moisture profiles of the clear-sky condition at a spatial resolution of 50 km. We have compared the 18- and 42-h predicted temperatures at lower (850 hPa), middle (500 hPa), and upper (200 hPa) levels with AIRS observations around (±1.5 h) 1800 UTC. The 1800 UTC AIRS observations are selected for the verification because of the greater spatial coverage of observations around this time. All the statistics shown in this section were obtained by comparing 30 samples of model forecasts with a corresponding sample of 30 AIRS observations. The spatial distribution of the biases and RMSEs in the predicted temperatures is computed for CNT to verify the temperature prediction of the WRF and the spatial distribution of FI is computed for the assimilation experiments to assess the impacts of satellite data. The spatial distribution of the biases in the 18-h temperature forecasts at 850 and 200 hPa from CNT show that the WRF produced a negative bias (warming) of the order 1–1.5 K, whereas at 500 hPa it showed a positive bias (cooling) of the same order in magnitude (not shown). The 18-h predicted temperature from WRF CNT showed RMSEs of the order of 1.5 K at 850 hPa (Fig. 8a), 1 K at 500 hPa (Fig. 9a), and 2 K at 200 hPa (not shown) with the higher values found over the northwest of Bay of Bengal and the Indian landmass. The spatial distribution of FI in 18-h temperature forecasts at 850 hPa showed that SSM/I TPW assimilation (SWV) produced the greatest improvement in temperature prediction, while the SSM/I wind speed (SW) showed slight degradation (Figs. 8b–e). Assimilation of satellite data showed improvement (more positive area as compared to negative area) in 18-h temperature forecasts at 500 (Figs. 9b–e) and 200 hPa (not shown). Analyses of 42-h temperature forecasts from different experiments at different pressure levels showed that assimilation of satellite data improved the temperature prediction and the largest improvement was due to QuikSCAT wind (QW) or SSM/I TPW (SWV) assimilation (not shown).
Since complete spatial coverage of AIRS observations over the model domain was not available for verification, we also compared the model-predicted temperature field with the NCEP analysis. The 24-h forecast from the WRF CNT simulation at 850 hPa showed RMSEs of the order similar to the comparison with AIRS, with larger RMSEs observed over the regions to the north and west of northwest India (Fig. 10a). Among the assimilation experiments, the largest improvements in the 24- (Figs. 10b–e) and 48-h (not shown) temperature forecasts at 850 hPa are observed to be due to the assimilation of SSM/I TPW (SWV), which is similar to the comparison with AIRS observations. The assimilation of QuikSCAT wind (QW) improved the 24- (Fig. 10b) and 48-h (not shown) temperature forecasts at 850 hPa over the southern part of the Arabian Sea. The experiments, in which SSM/I wind speed is assimilated (SW and SWWV), degraded the 24- (Figs. 10c and 10e) and 48-h (not shown) temperature forecasts at 850 hPa over most of the region. The 24-h temperature forecasts at 500 hPa by WRF CNT showed RMSEs of the order of 1 K and the largest values were distributed over the north and northwest of India (not shown). The assimilation of satellite data improved the 24- and 48-h temperature forecasts at 500 hPa with the largest improvement due to SSM/I TPW (SWV) assimilation (not shown).
3) Relative humidity
Similar to temperature, 18- and 42-h forecasts of relative humidity at different pressure levels are compared with AIRS observations valid at 1800 UTC. The 18- and 42-h forecasts by WRF CNT simulated a moist atmosphere (∼5%–10%) at 850 hPa as compared to AIRS, but underpredicted the humidity over the western part of the Bay of Bengal (not shown). Similarly, 18- and 42-h forecasts by WRF CNT overpredicted the humidity on the order of 10%–15% at 500 hPa (also not shown). The 18-h forecasts by WRF CNT showed a RMSE on the order of 10% at 850 hPa (Fig. 11a) and 20% at 500 hPa (Fig. 12a). The RMSE for the predicted humidity at 500 hPa over the Indian landmass is larger compared to the oceanic region (Fig. 12a). The RMSEs in 42-h forecasts by WRF CNT show a spatial distribution similar to the 18-h forecast with a slightly larger magnitude (not shown). The FIs due to the assimilation of satellite data are computed for those grid points where CNT RMSEs exceed a minimum threshold (10% at 850 hPa and 15% at 500 hPa). It is seen that the assimilation of satellite data improved the 18- (Figs. 11b–e) and 42-h (not shown) humidity forecasts at 850 hPa, and among the experiments, the largest improvement is observed to be due to SSM/I TPW (SWV) assimilation. Even though there are small pockets of negative area (degradation) observed in the 18- (Figs. 12b–e) and 42-h (not shown) humidity forecasts at 500 hPa from different assimilation experiments, a large positive area (improvement) is seen over the Indian landmass and the Bay of Bengal, particularly for the experiments in which SSM/I TPW is assimilated.
The 24- and 48-h forecasts of relative humidity at 850 and 500 hPa are also verified against the NCEP analysis. An RMSE of the order of 10% is observed in the 24-h predicted relative humidity (Fig. 13a) by WRF CNT at 850 hPa and, as expected, the RMSE is increased (∼5%) at 48 h (not shown). A large area of very high RMSE in humidity prediction is observed over the Arabian Sea in the 24- (Fig. 13a) and 48-h (not shown) forecasts. The assimilation of SSM/I TPW (SWV) improved the 24- (Fig. 13d) and 48-h (not shown) humidity forecasts, but the assimilation of wind speed (QuikSCAT or SSM/I) did not show much improvement (Figs. 13b and 13c). At 500 hPa, the RMSEs in the 24- and 48-h predicted humidity by WRF CNT are on the order of 20% (not shown). Similar to 850 hPa, assimilation of SSM/I TPW (SWV) showed significant improvement in humidity prediction at 500 hPa whereas wind speed (QuikSCAT or SSM/I) assimilation did not show much improvement (not shown).
The Tropical Rainfall Measuring Mission (TRMM) 3B42 version 6 (3B42V6) product (Haddad et al. 1997; Adler et al. 2000) was used for the validation of the model-predicted rainfall. Here, we examine the spatial distribution of the monthly accumulated rainfall from day 1 (accumulated rainfall during first 24-h forecast) and day 2 (accumulated rainfall from 24- to 48-h forecasts) forecasts, the improvement parameter (η), the statistical skill scores at various rainfall thresholds, and the impact ratio (IR) from the FIs based on the RMSEs in the rainfall forecasts.
The observed monthly accumulated rainfall distribution from TRMM (Fig. 14a) showed the July monsoon rainfall maxima over the west coast of India and the northeast Bay of Bengal near the Arakan coast. Much less precipitation is observed over semiarid regions of northwest India and the rainshadow areas of the eastern coast of southern peninsular India. The observed monthly accumulated rainfall maxima and rainshadow regions were qualitatively reproduced by accumulated rainfall prediction from WRF CNT day 1 forecasts (Fig. 14b). The difference plot (Fig. 14c) of the accumulated rainfall prediction from day 1 forecasts by WRF CNT with TRMM-observed rainfall showed that WRF CNT underpredicted rainfall over the Western Ghat near peninsular India and in the foothills of the Himalayas. The rainfall over the northern fringes of the Western Ghat and the head of the Bay of Bengal is overpredicted in WRF CNT day 1 forecasts. The accumulated rainfall from day 2 forecasts by WRF CNT showed an increased overprediction of rainfall throughout the model domain as compared to the day 1 forecast (not shown). For the quantitative assessment of improvement/degradation due to the assimilation of satellite data as compared to CNT, we have computed the spatial distribution of η in the monthly accumulated rainfall prediction from the day 1 and 2 forecasts. It is clear from the spatial distribution of η (Fig. 15) that the assimilation of SSM/I TPW (SWV and SWWV) improved (positive area in Figs. 15c and 15d) the day 1 rainfall prediction over the west coast of India and the Bay of Bengal. Similarly, the spatial distribution of η for the accumulated rainfall from day 2 forecasts also showed that the SSM/I TPW assimilation improved the rainfall prediction in this period (not shown). This impact is likely due to the improvement in the lower-level moisture fields due to the assimilation of SSM/I TPW.
To examine the skill of the different experiments in reproducing the frequencies of occurrence of rainfall events at or above a precipitation threshold, we have computed statistical skill scores [such as the bias scores (BSs) and equitable threat scores (ETSs); Anthes et al. (1989); Wilks (2006)] for 24-h accumulated rainfall predictions from day 1 and 2 forecasts. These statistics are obtained by comparing 30 samples each of daily accumulated rainfall predictions from different experiments with corresponding observed rainfall for the period 2–31 July 2006 at various rainfall thresholds (0.2, 0.5, 1, 2, 3, 4, 5, 6, 7, and 8 cm). A detailed description of the rain contingency table and formulas used for calculating the skill scores are available in Yang and Tung (2003). The variations of BSs and ETSs with rainfall thresholds obtained for day 1 and 2 forecasts are presented in Figs. 16 and 17, respectively. The day 1 forecasts from all the experiments overpredicted (underpredicted) the area of 24-h accumulated rainfall for low (high) thresholds (Fig. 16a). The assimilation of SSM/I TPW showed comparatively low bias (less overprediction) in predicting light (less than 2 cm) rainfall while it showed high bias (more underprediction) for heavy (more than 2 cm) rainfall (Fig. 16a). The comparatively higher bias (underprediction) due to SSM/I TPW assimilation for high rainfall thresholds may be due to the saturation problem in the SSM/I TPW data (Singh et al. 2008a). The day 2 forecasts by all the WRF experiments showed large overprediction (more than double that observed) in the area of occurrence of 24-h accumulated rainfall for all rainfall thresholds, and among the experiments, the overprediction is comparatively less in SWV (Fig. 16b). It can be seen in Fig. 17 that day 1 and 2 forecasts from the WRF showed poor skill in reproducing the frequency of occurrence of the daily accumulated rainfall at higher thresholds. Model skill deteriorated rapidly with rainfall threshold and the experiments in which SSM/I TPW is assimilated (SWV and SWWV) showed comparatively better skill in predicting the area of rainfall for most of the thresholds (Fig. 17). As expected, the skill in predicting the area of rainfall by WRF is poorer for the day 2 forecasts as compared to the day 1 forecasts (Figs. 16 and 17).
The BSs and ETSs based on the contingency table only measure model accuracy based on the frequency of occurrence at or above a precipitation threshold and, thus, do not determine the quantitative rainfall prediction skill of the model (Colle et al. 1999). Therefore, it is important to calculate the model skill based on RMSEs in the model forecasts using the actual quantitative precipitation from the model. To assess the percentage improvement–degradation due to the assimilation of satellite data as compared to CNT (without satellite data), we calculated the IR (see the appendix) based on quantitative FIs (see the appendix) for day 1 and 2 forecasted rainfall by comparing with the TRMM observed rainfall. The IRs of different assimilation experiments against the FI range for daily accumulated rainfall prediction from day 1 and 2 forecasts are shown in Fig. 18. Assimilation of satellite data (QuikSCAT and SSM/I TPW) produced consistent positive impact (IR values greater than 100 for high FI ranges) on rainfall prediction (Fig. 18), while SSM/I wind speed assimilation shows mixed (negative and positive) results. The assimilation of QuikSCAT winds yielded the maximum positive impact for day 1 forecasts. This impact likely occurs through a change of the boundary layer processes (e.g., moisture flux and convergence fields; because QuikSCAT contains wind speed as well as direction). Among the experiments, the assimilation of SSM/I TPW produced highest positive impacts in the day 2 rainfall predictions (Fig. 18b).
5. Sensitivity of assimilation results to the cumulus parameterization
The sensitivities of model biases and RMSEs to the convective parameterization scheme (CPS) used (KF scheme is used in this study) are explored by conducting some additional experiments. We have repeated a portion of the study (for the first week of July) using another CPS, namely the Grell–Devenyi ensemble (GDE; Grell and Devenyi 2002), in WRF. The results from these experiments are compared with the previous results that used the KF scheme. Such a comparison can answer two questions: 1) to what extend can the model biases and RMSEs described in this paper be attributed to the CPS used in the study and 2) how sensitive are the assimilation results to the CPS used. A comparison of model biases in 18-h temperature predictions by WRF CNT at various levels using the KF and GDE CPSs is presented in Fig. 19. It is clear that the simulation using GDE is similar to that with KF, except for the warming seen in the midlevels. As compared to KF, the magnitude of the biases is slightly larger in experiments using the GDE scheme. As seen earlier in the KF case, the 18-h humidity prediction using the GDE scheme also showed a moist atmosphere as compared to observations in the lower and midlevels (not shown). Similar results are also observed for 42-h temperature and humidity predictions. The RMSEs in the WRF CNT simulations and the FIs from different satellite data assimilation experiments using the GDE scheme for 18-h temperature and relative humidity prediction at 850 hPa are shown in Figs. 20 and 21, respectively. The WRF CNT simulations using the GDE scheme showed RMSEs of about 1.5 K (Fig. 20a) in the 18-h temperature prediction at 850 hPa, which is similar to the result from the experiments using the KF scheme (Fig. 8a) in this study. Similarly, the magnitudes of the RMSEs in the 18-h temperature prediction at mid- and upper levels by WRF CNT using the GDE scheme (not shown) are similar to those obtained using the KF scheme earlier in this study. The impacts of the satellite data on the WRF experiments with the GDE scheme (Figs. 20b–e) are in agreement with the simulations using the KF scheme (Figs. 8b–e). In both cases, the largest positive impacts are a result of the SSM/I TPW assimilation and degradation due to the SSM/I wind speed assimilation. Similarly, the largest positive impacts in the 18-h humidity prediction at 850 hPa in the experiments using the GDE scheme are observed to be due to SSM/I TPW assimilation while the least positive impacts are seen in SSM/I wind speed assimilation (Figs. 21b–e) and are in agreement with the earlier results using the KF scheme (Figs. 11b–e).
It is seen from the spatial distribution of the RMSEs in temperature and humidity prediction using the KF and GR schemes that the former showed comparatively isolated and extreme values of RMSEs as compared to the latter. One of the reasons for isolated and extreme values of RMSEs in experiments using the GR scheme may be due to the false precipitation rate simulated by this scheme, particularly the convective precipitation rates. The GR scheme had less active convective precipitation and might have failed in stabilizing the atmosphere through removing the instability locally by the transfer of heat and moisture. Similar results were also reported by previous investigators (Wang and Seaman 1997; Ratnam and Cox 2006). Even though slight differences are found in the satellite data assimilation experiments using different CPSs, on average the satellite data produced similar impacts to the WRF simulations in this study.
A month-long series of numerical simulations was conducted using the Advanced Research version of the Weather Research and Forecasting (WRF-ARW) model to assess the impacts of the assimilation of satellite data on model forecasts over the Indian region during the 2006 summer monsoon. The WRF and its 3DVAR system are used to investigate the impacts of the QuikSCAT near-surface winds, SSM/I-derived winds, and total precipitable water (TPW). The control (without satellite data assimilation) and experimental (with satellite data assimilation) forecasts are made for 48 h each day starting at 0000 UTC from 1 to 31 July 2006. The control run served as a baseline for verifying the assimilation experiments.
The 24- and 48-h forecasts from the WRF CNT simulations showed a weakening of the cross-equatorial flow over the southern Arabian Sea near peninsular India. The WRF CNT simulated a warm and moist atmosphere at the lower (850 hPa) levels, a cooling and moistening at the middle (500 hPa) levels, and a warming at the upper (200 hPa) levels. The 18–24-h forecasts at lower levels from WRF CNT showed RMSEs in predicted wind, temperature, and humidity to be on the order of 2 m s−1, 1.5 K, and 10%–15%, respectively, with the magnitude of the RMSEs being slightly larger in the 42–48-h forecasts. The errors in the wind speed predictions of the WRF are larger over the Bay of Bengal as compared to the Arabian Sea. Assimilation of satellite data (QuikSCAT wind and SSM/I TPW) improved the 24- and 48-h hour predicted wind speeds, with the largest improvement observed due to SSM/I TPW assimilation. Among the assimilation experiments, the largest positive impacts in the temperature and humidity predictions are observed to be due to SSM/I TPW assimilation. The assimilation of SSM/I wind speeds resulted in a degradation of the temperature and humidity predictions at lower levels.
The day 1 forecasts by WRF CNT showed an underprediction of rainfall over the Western Ghat near peninsular India, as well as over the foothills of the Himalayas. The spatial distribution of the improvement parameter (η) from the day 1 and 2 forecasts showed significant improvement in the rainfall prediction over the west coast of India due to the assimilation of SSM/I TPW. The day 1 forecasts of WRF overpredicted (underpredicted) the area of light (heavy) rainfall in the 24-h accumulated rainfall predictions. The experiments in which SSM/I TPW is assimilated showed comparatively better skill in predicting the areas of rainfall for most of the thresholds tested. The assimilation of QuikSCAT winds significantly improved the quantitative rainfall prediction skill level for the day 1 forecasts, but significant improvement due to QuikSCAT data was not observed in the day 2 rainfall predictions. The SSM/I TPW assimilation resulted in improved rainfall prediction skill for both day 1 and 2 forecasts. Across most of the thresholds, the assimilation of SSM/I wind speeds alone degraded the day 1 and 2 rainfall prediction skill of the WRF.
WRF is made publicly available and supported by the Mesoscale and Microscale Meteorology (MMM) division at the National Center for Atmospheric Research (NCAR). Their dedication and hard work is gratefully acknowledged. The authors acknowledge the National Centers for Environmental Prediction (NCEP) for making analysis data available at their site. The SSM/I and QuikSCAT data were obtained online (ftp.ssmi.com). The ICOADS data were also obtained online (ftp.dss.ucar.edu). The AIRS and TRMM data were obtained from NASA Web sites and NASA is acknowledged thankfully. The authors thank the anonymous reviewers for their critical and insightful comments and suggestions, which were helpful in substantially improving the content and quality of the manuscript. One of the authors (RV) acknowledges the award of research fellowship by the University Grant Commission (UGC), India.
Definition of the Parameters Used for Assessing the Impacts of Satellite Data
- The root-mean-square sensitivity (RS) of the assimilated satellite data in the 3DVAR analysis (initial conditions) is defined as (A1) only contains two analyses and does not contain verification with an independent observation, the sensitivity diagnosed by Eq. (A1) does not indicate whether the initial conditions from the satellite data assimilation are better or worse compared with the control (without satellite data). The purpose of such a sensitivity analysis is to find out how much and where the satellite data assimilation impacted the initial analysis.
- Forecast impact (FI) based on the RMSE in the model forecasts is defined following Wilks (2006) as (A1). The procedure of dividing the forecast error by error in the control forecast and multiplying by 100 normalizes the results and gives a percentage improvement in the predicted meteorological parameters with respect to the control forecast irrespective of its original magnitude. A positive FI means the forecast with the assimilation of satellite data compares more favorably with the observation than the control forecast (without satellite data).
Corresponding author address: Rakesh V., CSIR Center for Mathematical Modeling and Computer Simulation, NAL Belur Campus, Bangalore 560037, India. Email: firstname.lastname@example.org