1. Introduction
Flash floods and wild fires are the two leading weather-related natural disasters in Southern California. Of the 10 highest 24-h rainfall totals in California, all 10 have occurred in Southern California during winter. Heavy rainfall events, in conjunction with steep terrain slopes and the formation of hydrophobic soil layers due to wild fires near the beginning of the wet season, enhance the chances of flash floods in Southern California. This has made quantitative precipitation forecast (QPF) and streamflow forecasting among the most important tasks for the National Weather Service Office (NWSO)–Los Angeles and Ventura County Watershed Protection District (VCWPD), respectively, for supporting the emergency response agencies in the region.
The QPFs and streamflow forecasts at these agencies have been experiencing difficulties due to a lack of fine-resolution precipitation data needed to resolve significant spatial structures in rainfall associated with extreme variations in local terrain. In an attempt to improve flood forecasting, in particular finescale QPFs and streamflow forecasts for the mountainous Santa Barbara and Ventura County watersheds, the NWSO and VCWPD, in collaboration with the authors at the University of California, Los Angeles, (UCLA) and the Jet Propulsion Laboratory/California Institute of Technology (JPL) have developed the Integrated Regional Forecast System (IRFS). The IRFS is composed of a regional numerical weather prediction (NWP) model running typically at sub-10-km resolution, and a watershed-hydrology model nested within operational NWP data provided by the National Centers for Environmental Prediction (NCEP). Initial tests of the system showed that one of the keys for improving short-term QPF is accurate initial atmospheric water vapor fields for the regional NWP model (Marcus et al. 2004). The work described here is directed toward application of GPS precipitable water vapor (PWV) retrievals to the IRFS.
For the regional NWP produced by the IRFS, the water vapor fields as well as other atmospheric variables needed for initialization and time-dependent lateral boundary conditions are obtained from operational 40-km-resolution Eta forecast data at NCEP. Water vapor fields in these data, both initial analyses and forecasts, often suffer from inaccuracies due to a lack of local observations and incomplete NWP model physics, especially over oceans where the water vapor fields are almost entirely generated by the model. It has been found that, although the fine spatial resolution of the IRFS generally improves QPFs in Southern California compared to those based on operational NWP data used to drive the regional forecasts, the amount of water vapor in operational NWP data is often underestimated during heavy rainfall periods and in turn limits the QPF improvements by the IRFS (Marcus et al. 2005). Hence, accurate local water vapor observations, in addition to fine spatial NWP model resolution, are crucial for improving QPFs and streamflow forecasts for the region. Despite its importance, local water vapor data from conventional observations are severely limited because of a lack of radiosonde stations in the region, particularly over the adjacent Pacific Ocean.
The dense GPS array in Southern California, in particular the Southern California Integrated GPS Network (SCIGN), can provide high-quality atmospheric water vapor data necessary for improving the accuracy of the initial water vapor fields for the IRFS-based regional forecasts. The phase delay of L-band (∼19 cm) radio signals, propagating from the GPS satellite constellation to ground-based receivers, provides information on the water vapor content of the intervening atmosphere. Since the GPS wavelength is much longer than typical drop sizes, the PWV retrieval is not significantly affected by the presence of clouds and/or rain, making it ideal for the investigation of water vapor variations and the prediction of rainfall during storm events. De Pondeca and Zou (2001) found that inclusion of GPS zenith delay data into model initialization showed a positive, although small, impact on rainfall forecasts in Los Angeles County. Gutman et al. (2004) report that addition of GPS PWV resulted in improvements in 3-h relative humidity forecasts for a part of the central United States, especially during the cold season, despite the fact that this region is one of the most conventional data-rich regions in the world. Kuo (2006) showed that GPS PWV is useful for improving atmospheric water vapor analysis and short-term precipitation forecasts. Nakamura et al. (2004), however, found that inclusion of PWV retrievals from the dense GPS network in Japan had an almost neutral impact on rainfall forecasts. Taken together, these studies show that the impact of additional GPS-retrieved water vapor data on NWP varies according to region and season. Hence, the application of GPS PWV to a regional forecast system needs careful benchmarking for each region to which the system is applied.
In what follows, section 2 describes the datasets, models and procedures used to evaluate the effect of GPS retrievals on QPF in Southern California, section 3 presents results from case studies of GPS impact on QPF during heavy precipitation events in February 1998 and January 2005, and statistics from an ensemble of 47 winter storm cases spanning February 1998–January 2005 are analyzed in section 4. Section 5 provides a discussion of our results and conclusions from our study.
2. Experimental design
The Southern California basin is extensively covered by stationary ground sites that track GPS–constellation radio signals continuously in time. The transmission delay between the satellites and a ground station is sensitive to the water vapor content along the ray path (e.g., Foster et al. 2000). Typically, the transmission delay at each ground station is reported as a total zenith delay (TZD), or the effective delay directly overhead from the station, which is computed by combination of a number (at least four) of delay values along slant ray paths assuming isotropic atmospheric conditions (Bar-Sever et al. 1998). The advantages of using TZD to estimate PWV include its all-weather capability, high temporal resolution (approximately 5 min, with some data voids), its concentration over land areas where other satellite water vapor retrievals are difficult, and remarkable stability of the retrieved phase delays to sensor drift and calibration (Haines and Bar-Sever 1998). SCIGN covers the region of our interest, and stations in other GPS networks, such as the International GNSS Service (IGS), supplement these. By November 2006 there were over 300 regional GPS stations as a result of recent rapid growth in these networks (Fig. 1; updated maps can be viewed online at http://sopac.ucsd.edu/maps). During our study period the number of regional GPS sites was more limited, growing from fewer than 50 stations in February 1998, to about 225 in March 2001; in January 2005 nearly 250 stations were available to provide TZD time series.
a. Derivation of PWV from GPS zenith delay
To compute PWV from the TZD, surface pressure and temperature are needed to separate hydrostatic delay (due to the total mass of the atmospheric column) and wet delay (due to the permanent dipole moment of the moisture within the column) from the TZD. The estimated hydrostatic delay is subtracted from the TZD data to yield the wet delay, which is nearly proportional to PWV (Bevis et al. 1992). There are only a limited number of GPS sites equipped with “met packs” (e.g., on the order of 10 sites within the SCIGN) that can provide surface pressure and temperature data. We have instead estimated these surface values using the 40-km Eta model initial analysis by cubic spline fitting over the horizontal and exponential interpolation along the vertical axes, respectively. Where the met pack data are available, we have found reasonably close agreement between the hydrostatic delay values obtained using the Eta analysis and the values computed using the met pack measurements. Specifically, the PWV value is more sensitive to the surface pressure than temperature, where the former needs to be evaluated with an accuracy of 5 mb to give 1-cm accuracy in the wet delay that varies typically by 10–30 cm during a frontal passage (Bevis et al. 1992). Our comparison of Eta-based surface pressure values against the met pack counterparts shows a mean bias of less than 2 mb and a root-mean-square (RMS) error of less than 1 mb.
Figure 2 compares the GPS PWV time series at two near-coastal stations in our region of interest (the Harvest oil platform off Cape Conception and the city of Long Beach) against coincident PWV values interpolated spatially from the 40-km Eta analysis during the first week of February 1998. It shows that a key advantage of the GPS data is their shorter sampling period (15 min) that can resolve finer-scale features of the water vapor field that are not captured by the 12-hourly Eta data; in addition, Eta analyses often underestimate PWV substantially during heavy rainfall periods.
b. Model initialization procedures
The PWV retrievals thus obtained over the GPS stations are used to update the Eta PWV values; to maximize efficiency for operational applications a standard 2D objective analysis technique is used (Barnes 1964). The gridded Eta PWV values are first interpolated horizontally to the location of each GPS receiver by cubic spline, and a PWV anomaly is computed at that location. Since the vertical water vapor distribution is not known a priori, there is no convenient way to systematically scale the PWV with altitude between nearby grid points, so we did not include an elevation correction in the grid–GPS interpolation of the PWV values. In regions of elevated topography, the modeled and observed PWV, and hence its anomaly, will be relatively small, so we do not expect such locations to have a dominant effect on our analysis. In the second step, the anomaly PWV values computed at the GPS sites are horizontally interpolated back to the grid locations using a Gaussian-weighted influence function with an e-folding length scale of 50 km; this is somewhat larger than the typical separation of GPS receivers in the SCIGN array, and was found to produce PWV fields with an appropriate degree of smoothness for model initialization when added to the Eta-specified background. In the final step, the updated PWV values are distributed along the vertical grid points using the method of Kuo et al. (1993), which scales the background water vapor profile to match the vertically integrated amount to the given PWV value.
In the control experiments (CTL hereinafter), the initial and boundary data are obtained from the 40-km Eta analysis. To evaluate the impact of GPS PWV data on the regional QPF, we also conduct sensitivity experiments (GPS hereafter) where the 3D water vapor field in the initial condition is updated by the GPS TZD data as outlined above. Note that the only difference between the two experiments is their corresponding initial water vapor field. As Fig. 2 shows, there are many instances where the GPS and Eta PWV values are fairly close; thus, a given station value, by chance, may have little contribution in differentiating the CTL and GPS experiments.
c. Model description
The Mesoscale Atmospheric Simulation (MAS) model (Soong and Kim 1996; Kim 2005) interactively coupled with the NCEP–Oregon State University–Air Force Weather Agency–Office of Hydrology (Noah) land surface scheme (Kim and Ek 1995; Chang et al. 1999) is used in this study. The MAS model is a primitive equation, limited-area atmospheric model written on the σ coordinates in the vertical (Soong and Kim 1996). The advection equations are solved using the finite-difference scheme of Takacs (1985) that is characterized by minimal phase errors and numerical dispersion. A 4-class version of the bulk microphysics scheme of Cho et al. (1989) has been used to compute the precipitation in this study. Atmospheric radiative transfer is computed using the δ-2/4-stream Fu–Liou scheme (Fu and Liou 1993; Gu et al. 2003). A 4-layer version of the Noah land surface model (Kim and Ek 1995) is coupled with MAS to compute the land surface processes. Noah predicts the soil moisture content and soil temperature within model soil layers, as well as canopy-water content and snow-water equivalence. The temperature and specific humidity for calculating the surface sensible and latent heat fluxes, outgoing longwave radiation, and ground heat fluxes are calculated by iteratively solving a nonlinear form of the surface energy balance equation. For more details of the MAS and Noah models, see Mahrt and Pan (1984), Pan and Mahrt (1987), Kim and Ek (1995), and Soong and Kim (1996).
The experimental domain (shown in Fig. 3) covers Southern California with a 50 × 40 grid nest at a 9-km resolution in the horizontal, and 24 atmospheric and 4 soil layers in the vertical. The control (CTL) simulations are initialized using only the NCEP Eta initial analysis fields at 40-km horizontal resolution. The sensitivity (GPS) runs are initialized in the same way as the CTL run except that the initial water vapor fields are obtained by combining the PWV data from the SCIGN GPS array with that from the NCEP Eta initial data as described above. Both sets of simulations are driven by the same lateral boundary forcing obtained from the Eta initial fields. Hence, the CTL and GPS runs differ only in their initial water vapor fields.
3. Case studies
a. El Niño storm of February 1998
On 2–3 February 1998, one of the strongest storms of the “El Niño” winter impacted the Southern California coast, causing extensive damage because of high winds and flooding, which resulted in landslides and coastal erosion. Figure 3 shows the initial PWV distribution at 0000 UTC 2 February 1998, as interpolated from the Eta initial analysis at the corresponding time. Note that the Eta water vapor field is characterized by weak horizontal gradients over the ocean, indicating a lack of detailed information on water vapor fluctuations offshore. As described by Neiman et al. (2004), however, the approaching storm had a rich offshore mesoscale frontal structure, with alongshore moisture gradients (see below) not detected by the Eta analysis. Here we examine the impact of GPS data, used to estimate PWV amounts at receivers of the SCIGN array, on short-term (up to 12 h) forecasts of rainfall amounts on 2 February 1998. We also investigate the period over which the QPF is influenced by the additional GPS PWV data, using airflow trajectory analysis from the model runs.
The filled circles in Fig. 4a represent the locations of the GPS receivers used for the case study forecast, with the PWV value for 0000 UTC 2 February 1998 indicated by the color scale. A strong northwest–southeast gradient in the offshore column water vapor is evident, which is not captured by the Eta analysis. Over the land, the areas of lowest PWV coincide with elevated topography, reflecting the higher concentration of water vapor in the planetary boundary layer (PBL). Figure 4b shows how the increased column values of water vapor at Point Conception (northernmost coastal station in Fig. 4a) are vertically distributed in the model. The highest concentrations are found in the PBL, extending to an altitude of about 600 m.
Figure 5 shows (in both panels) the change in the initial PWV distribution at 0000 UTC 2 February 1998 due to the addition of the GPS PWV data. A positive anomaly appears around Point Conception, extending along the coast and reaching its greatest southward extent in the vicinity of the San Nicolas Island receiver. Some enhancement in water vapor is also seen over the more heavily instrumented area of the SCIGN receivers in the Los Angeles basin. Hatched lines show the simulated trajectories of neutrally buoyant tracers released at 300 m and 700 m above sea level (upper and lower panels, respectively) at 0000 UTC 2 February 1998, with cross marks indicating the positions reached at half-hourly intervals. As is typical for Southern California rain events the trajectories move onshore from a southerly or southwesterly direction (Nezlin and Stein 2005), so that information from coastal and/or island stations is critical for improving knowledge of upstream water vapor conditions.
Figure 6 shows the increase in the forecasted 6-h cumulative precipitation due to the incorporation of GPS data, in relation to the verification boxes situated over Santa Barbara (SB; left-hand box) and Ventura County (VC; right-hand box). Both verification areas are located on the southern slopes of the Transverse Range (west–east zone of elevated topography seen in Fig. 4a), and hence are vulnerable to flooding due to local orographic precipitation, especially during the southerly flow regime that characterized the case study period. The largest simulated QPF increase is located over the SB box, reflecting the large PWV anomalies detected in the vicinity of nearby Point Conception. Because of the impact of the PWV data, the GPS predicted cumulative precipitation in the SB box closely follows the observed for the first few hours after initialization, while the CTL run largely fails to predict the observed SB rainfall (Fig. 7a). Because of the rapid advection of the GPS-detected water vapor through the coastal zone (Fig. 5), however, the beneficial effects of the additional GPS data last for only a few hours, with subsequent large amounts of observed precipitation not forecasted in the 0000 UTC-initialized runs.
For the more easterly VC box, which is less directly influenced by the large PWV anomaly at Point Conception, the proportionate QPF differences between the GPS and CTL forecasts are not as large, although the GPS QPF tracks the observed cumulative precipitation quite closely for the first four hours (Fig. 7b). As for the SB forecast, however, the subsequent large observed precipitation for the VC box is not captured by the model runs. Nevertheless, it is interesting to note that the QPF enhancement for the GPS forecast continues to increase for up to 6 h after initialization, reflecting the greater southerly extent of the GPS-detected PWV anomaly in the vicinity of the San Nicolas island receiver, located almost directly upstream of the VC box in the 300-m wind field (Fig. 5).
Thus while the GPS-sensed PWV improved the precipitation forecast for the 0000 UTC 2 February 1998 case, the benefits are of a limited duration since the forecast time span affected by the GPS data is restricted by the spatial coverage of the receivers in the region. The low-level particle trajectory analysis (Fig. 5) shows that the initial water vapor fields affected by the additional PWV data pass through the SB and VC forecast areas in 3–6 h from the initial time. Because of the onshore trajectories that prevailed during this event, receivers located over coastal and island stations were the principal means for detecting the upstream PWV anomalies that led to the successful forecast improvements. As will be seen in the following, however, additional water vapor data over the eastern Pacific region, incorporating more extensive island-based receiver network and/or satellite retrievals, will be crucial for increasing the useful forecast lead time over the heavily populated coastal strip of Southern California.
b. La Conchita storms of 7–10 January 2005
During the July 2004–June 2005 rainfall season, Southern California experienced one of its wettest winters on record, with Los Angeles recording its second-highest annual rainfall (37.25 inches compared with an average of 15.14 inches) and over 8 inches of rain in each of the consecutive months December 2004–February 2005. On 10 January 2005, after several days of local heavy rainfall (Fig. 8), the small coastal community of La Conchita, California, located between the SB and VC boxes, experienced a shallow but fast-moving landslide that destroyed several homes and caused 10 fatalities. During the three days preceding the landslide, both the SB and VC 3- and 6-h QPFs showed substantial improvements with addition of GPS PWV data (Fig. 9—note that this is not a continuous record, since the QPFs were initialized only at 12-hourly intervals). The hourly precipitation for La Conchita, taken from Eta-only and Eta + GPS forecasts initialized at 1200 UTC 10 January 2005, are compared with observed hourly totals in Fig. 10. The GPS-initialized forecast produces an improved hourly precipitation record when compared with the CTL case, although both QPFs far underestimate the observed rainfall amounts (divided here by a factor of 5 for visualization purposes). It is also noteworthy that the GPS QPF maximizes 1 h earlier than both the observed precipitation and the Eta-only forecast, indicating that the impact of GPS-derived PWV on the simulated rainfall was too brief to capture the magnitude or timing of the 10 January storm that may have been the immediate trigger for the landslide (Jibson 2005). The results of this comparison serve to underscore the conclusion from the El Niño case study, which showed that the addition of GPS PWV data from coastal/island stations can impact regional QPFs in the coastal zone for only a few hours.
4. Evaluation of ensemble statistics
To test the robustness of the conclusions reached for the case studies presented above, we compare the results of an ensemble of 47 GPS and CTL forecasts, made for winter storm events during the interval February 1998–January 2005, for the SB and VC regions. Each forecast was initialized at either 0000 UTC or 1200 UTC, with the cumulative 3- and 6-h QPFs archived for comparison with concurrent observations, thus allowing for some degree of statistical independence between successive forecasts. The cases examined are shown in Table 1 (these include the case studies mentioned above).
The mean QPFs and corresponding observations and bias estimates are given in the Table 2. For the winter storm cases chosen for this study the model forecasts capture an average of only about 30% (40%) of the SB (VC) observed precipitation, although the QPFs overestimated rainfall in a significant fraction (22%) of the cases. For the 3-h forecasts the GPS retrievals increase the predicted rainfall by about 8% (6%) for the SB (VC) region, while for the 6-h forecasts the average GPS-initialized rainfall shows a slight decrease of about 1% (3%) from the CTL QPFs for these regions.
To separate the effects of the low bias of the model QPFs from that of adding GPS PWV retrievals to the initial water vapor fields, we investigated the variance explained by both CTL and GPS forecasts, allowing for an arbitrary QPF gain factor for each category of forecast. The mean QPF for each category, as well as for the observed precipitation, was subtracted from the data before performing the calculation. The results (Fig. 11) show that, for all four categories considered, the GPS forecasts explain more of the observed precipitation variance than the CTL cases for a unit scale factor (i.e., no gain applied), and also explain a greater maximum variance for arbitrary gain. Similar results are obtained when the means were not subtracted prior to computing variance, although the maximum explained variance then occurs for higher amplification factors (not shown). These results show that the GPS retrievals provide robust albeit moderate benefits to the 3- and 6-h QPFs for the SB and VC regions, reflecting added skill that is not simply due to changes in the mean level of forecasted precipitation.
For the 40 cases in which the CTL forecast overestimated the observed precipitation, in fact, the GPS-initialized QPF was more accurate 25 times (62.5%), while for the 138 cases in which the CTL forecast underestimated the observed precipitation, GPS was more accurate 78 times (56.5%; note that cases with equal CTL and GPS QPFs were omitted from this calculation). Thus, the GPS forecasts exhibit skill in decreasing overpredicted precipitation, as well as increasing underpredicted amounts.
To examine the skill of the GPS forecasts in more detail, we computed the improvement (i.e., the reduction in the absolute QPF error) due to incorporation of GPS data for each of the 47 cases, with no amplification factor applied (upper portion of Table 3). The normalized improvements, defined as the average improvement divided by the standard deviation (σ) as estimated from the 47 values for each category, are shown in the last column of Table 3. Note that all 4 of the regional categories show positive improvements for the GPS-initialized cases, although the average effect is small, with the 3-h forecasts generally showing a larger GPS effect. Combining the QPFs from the 2 regions into 94-case ensembles (middle portion of Table 3) gives the same average improvement; however the larger number of cases in each ensemble leads to higher statistical confidence, with the combined 3-h QPFs showing improvement significant at the 92.5% (1.83σ) level.
A subjective examination of the wind fields for the 47 cases showed that the largest improvements were generally found for PBL (300 m) wind directions from the southern quadrant (i.e., between SW and SE). For the 26 SB and VC cases (a total of 52) that met this criterion, the average 3- and 6-h QPF improvements were larger than those for the unrestricted ensemble by factors of 1.9 and 3.6, respectively, with the 6-h QPF improvement actually exceeding the 3-h value for these cases (lower portion of Table 3). As seen for the case study wind fields shown in Fig. 5, forecasts for southerly flow may benefit from a larger upstream area of GPS PWV retrievals provided by the island receivers (in particular San Nicolas), which can significantly lengthen the useful QPF lead time for the SB–VC area. It is also noteworthy that in spite of the smaller number of cases, the southerly flow ensemble showed average 3- and 6-h QPF improvements that were statistically significant above or near the 2σ level.
Figure 12 shows histograms of GPS improvements for the combined 94-case ensemble (left panel) and for the 52-case southerly wind subensemble (right panel). Although the average QPF improvement is less than 1 mm (see Table 3), a substantial number of cases show improvements of over 5 mm. For the full 94-case ensemble the 6-h GPS-initialized QPFs show increasing divergence from the 3-h forecasts (seen as a larger horizontal spread and lower maximum value of the dashed relative to the solid line), as initial differences are magnified by nonlinear effects. For the 52-case southerly ensemble, however, this “diffusion” effect is smaller, as the greater upstream extent of the receiver array provides longer lead time and higher predictability for the GPS-initialized QPFs.
Figure 13 shows the normalized improvement in QPF skill due to incorporation of the GPS data, for forecast durations of 1–6 h. For the individual regions statistical significance is achieved only for the first hour QPF, which is subject to spinup problems in both the CTL and GPS runs (see, e.g., Fig. 7). As noted above, the combined ensemble of 94 cases maintains statistically significant improvement for nearly the first 3 h, with average skill dropping to low levels (less than 1 σ) after 4 h. However, the QPF improvements for the combined subensemble of 52 cases with southerly winds at 300 m maintains normalized improvements near the 2σ level for the first 6 h, with the average skill level dropping sharply for longer forecasts (not shown). These results emphasize the importance of PWV retrievals from the island GPS stations, in particular San Nicolas, which provides data nearly 150 km upstream of the verification areas, for the southerly flows that bring moist tropical air into the SB–VC region during heavy rainfall events.
5. Discussion and conclusions
This study is intended to evaluate the potential for GPS-retrieved PWV to enhance the QPF capability in Southern California, particularly as related to natural hazard mitigation associated with flash flooding and landslides in the Santa Barbara (SB) and Ventura County (VC) areas targeted primarily by the IRFS-based NWP. Our results show that GPS does beneficially impact model-generated QPFs, with a greater fraction of the variance of observed precipitation explained for all four of the major forecast categories considered (SB and VC regions, for 3- and 6-h QPF), compared to CTL forecasts run with Eta-only initial water vapor conditions. On average the QPF benefits for the SB/VC area are rather small (less than 1 mm), although the additional GPS PWV data resulted in notable improvements during some heavy rainfall events, such as the El Niño storm of 2 February 1998 and the 7–10 January 2005 storms leading up to the La Conchita landslide. For many of the winter storm cases selected for this study, the model runs capture only a fraction (less than 40%) of the observed rainfall, emphasizing the difficulty of quantitative precipitation forecasting in regions of complex terrain (e.g., Roe 2005). It was interesting to note, however, that GPS-initialized forecasts showed skill in decreasing overpredicted precipitation (improving 25 out of 40 such cases) as well as increasing underpredicted amounts (78 out of 138 cases, not counting equal amounts).
For the El Niño (2 February 1998) storm that we examined in more detail, the combination of strong PWV anomalies and a southwesterly flow field led to GPS QPF improvement of several millimeters for both the SB and VC regions, for the first 4 h following initialization at 0000 UTC 2 February 1998. A trajectory analysis showed that GPS retrievals from coastal and island stations provide most of the mesoscale PWV data that enhances the model QPFs. However, since these stations are located within 150 km of the verification areas, the lead time during which the GPS data can influence the forecast is restricted to about 6 h or less. A similar situation prevailed for model forecasts for the La Conchita area (located on the coast between the SB and VC boxes) initialized at 1200 UTC 10 January 2005, a few hours in advance of the heavy precipitation that appears to have triggered the landslide that occurred later that day (Jibson 2005). Although the GPS QPF captured slightly more rainfall than the CTL forecast, the differences lasted for only a few hours after initialization, with the GPS rainfall peaking an hour before both the CTL QPF and the observed precipitation. The results of these case studies show that while GPS data can substantially benefit regional QPFs, more extensive PWV retrievals from a broader upstream area are required to fully capture the timing and magnitude of strong rainfall events in the SB–VC region.
The overall forecast skill, defined as the average absolute reduction in QPF error with the incorporation of GPS data, is positive for all four 47-case regional forecast categories considered; by combining the 3-h SB and VC QPFs into one 94-case ensemble, improvement significant at the 92.5% (1.83σ level) was attained. For the 52-case subensemble featuring southerly quadrant winds at the 300-m level the benefits are more robust, with both the average 3- and 6-h GPS QPF improvements being statistically significant above or near the 2σ level. The improved results for the southerly flow cases indicate that the GPS island receivers provide valuable PWV data for regional QPF associated with strong moisture advection from the subtropics, which improves the model NWP for orographic precipitation in the heavily populated coastal strip located on the southern flank of the west–east Transverse Range in Santa Barbara and Ventura Counties. The useful forecast lead time, however, is limited to no more than 6 h, since the furthest offshore receiver (San Nicolas) is located less than 150 km from the coastline. Clearly, the addition of PWV retrievals from satellite sensors, such as Atmospheric Infrared Sounder (AIRS) and Advanced Microwave Scanning Radiometer for Earth Observing System (AMSR-E) onboard the Aqua platform, has the potential to provide water vapor data further upstream over the eastern Pacific and thus lengthen the useful QPF period. The combination of the satellite data, which cover a broader upstream area but are available only twice a day, with the more spatially limited but temporally intensive GPS PWV retrievals, offers a good opportunity to maximize benefits from complementary sources of data. The forthcoming availability of GPS TZD from the California Real Time Network (CRTN, a subset of SCIGN; Genrich and Bock 2006) in near–real time, in conjunction with satellite-observed water vapor data, can help to realize these benefits for operational systems in the not-too-distant future.
Acknowledgments
Jean Dickey (JPL) and Yehuda Bock (SIO) provided valuable discussion and comments on our work. We thank Art Henkel, Rob Hartman, and Paul Neiman (NOAA) for sending us the ALERT rain gauge data used in this study and two anonymous reviewers for comments that helped to improve the manuscript. Support for this project was provided by the NASA SENH program (NAG5-13248); JK also received support from the NOAA GAPP program (NA03OAR4310012). Work by SM and TC was performed at the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration.
REFERENCES
Barnes, S. L., 1964: A technique for maximizing details in numerical weather map analysis. J. Appl. Meteor., 3 , 396–409.
Bar-Sever, Y. E., P. M. Kroger, and J. A. Borjesson, 1998: Estimating horizontal gradients of tropospheric path delay with a single GPS receiver. J. Geophys. Res., 103 , 5019–5035.
Bevis, M., S. Businger, T. A. Herring, C. Rocken, R. A. Anthes, and R. H. Ware, 1992: GPS meteorology: Remote sensing of atmospheric water vapor using the Global Positioning System. J. Geophys. Res., 97 , 15787–15801.
Chang, S., D. Hahn, C. Yang, D. Norquist, and M. Ek, 1999: Validation study of the CAPS model land surface scheme using the 1987 Cabauw/PILPS dataset. J. Appl. Meteor., 38 , 405–422.
Cho, H., M. Niewiadomski, and J. Iribarne, 1989: A model of the effect of cumulus clouds on the redistribution and transformation of pollutants. J. Geophys. Res., 94 , 12895–12910.
De Pondeca, M., and X. Zou, 2001: Moisture retrievals from simulated zenith delay “observations” and their impact on short-range precipitation forecasts. Tellus, 53A , 192–214.
Foster, J., and Coauthors, 2000: El Niño, water vapor, and the Global Positioning System. Geophys. Res. Lett., 27 , 2697–2700.
Fu, Q., and K-N. Liou, 1993: Parameterization of the radiative properties of cirrus clouds. J. Atmos. Sci., 50 , 2008–2025.
Genrich, J. F., and Y. Bock, 2006: Instantaneous geodetic positioning with 10–50 Hz GPS measurements: Noise characteristics and implications for monitoring networks. J. Geophys. Res., 111 .B03403, doi:10.1029/2005JB003617.
Gu, Y., J. Fararra, K-N. Liou, and C. R. Mechoso, 2003: Parameterization of cloud–radiation processes in the UCLA general circulation model. J. Climate, 16 , 3357–3370.
Gutman, S. I., S. R. Sahm, S. G. Benjamin, B. E. Schwartz, K. L. Holub, J. Q. Stewart, and T. L. Smith, 2004: Rapid retrieval and assimilation of ground based GPS precipitable water observations at the NOAA forecast systems laboratory: Impact on weather forecasts. J. Meteor. Soc. Japan, 82 , 351–360.
Haines, B. J., and Y. E. Bar-Sever, 1998: Monitoring the TOPEX microwave radiometer with GPS: Stability of columnar water vapor measurements. Geophys. Res. Lett., 25 , 3564–3566.
Jibson, R. W., 2005: Landslide hazards at La Conchita, California. USGS Open-File Rep. 2005-1067, U.S. Department of the Interior, 12 pp.
Kim, J., 2005: A projection of the effects of the climate change induced by increased CO2 on extreme hydrologic events in the western U.S. Climatic Change, 68 , 153–168.
Kim, J., and M. Ek, 1995: A simulation of the surface energy budget and soil water content over the Hydrologic Atmospheric Pilot Experiments-Modelisation du Bilan Hydrique forest site. J. Geophys. Res., 100 , 20845–20854.
Kuo, Y., 2006: Assimilation of ground-based GPS data for short-range precipitation forecast. Preprints, Fourth Korea–US Joint Workshop on Mesoscale Observations, Data Assimilation, and Modeling for Severe Weather, Seoul, Korea, KOSEF and NSF/OISE, 38–41.
Kuo, Y., Y. Guo, and E. Westwater, 1993: Assimilation of precipitable water measurements into a mesoscale model. Mon. Wea. Rev., 121 , 1215–1238.
Mahrt, L., and H. Pan, 1984: A two-layer model of soil hydrology. Bound.-Layer Meteor., 29 , 1–20.
Marcus, S. L., J. Kim, T. M. Chin, J. O. Dickey, D. Danielson, C. Jacobson, and J. Laber, 2004: A regional NWP system for southern California: Applications of GPS-derived PWV retrievals. Eos, Trans. Amer. Geophys. Union, 85 .Fall Meeting Suppl. Abstract A53A-0853.
Marcus, S. L., J. Kim, T. M. Chin, D. Danielson, and J. Laber, 2005: Impact of GPS-derived PWV data on regional QPF in southern California. Eos, Trans. Amer. Geophys. Union, 86 .Fall Meeting Suppl. Abstract G13A-04.
Nakamura, H., K. Koizumi, and N. Mannoji, 2004: Data assimilation of GPS precipitable water vapor into the JMA mesoscale numerical weather prediction model and its impact on rainfall forecasts. J. Meteor. Soc. Japan, 82 , 441–452.
Neiman, P. J., P. O. G. Persson, F. M. Ralph, D. P. Jorgensen, A. B. White, and D. E. Kingsmill, 2004: Modification of fronts and precipitation caused by coastal blocking during an intense landfalling winter storm in southern California: Observations during CALJET. Mon. Wea. Rev., 132 , 242–273.
Nezlin, N. P., and E. D. Stein, 2005: Spatial and temporal patterns of remotely-sensed and field-measured rainfall in southern California. Remote Sens. Environ., 96 , 228–245.
Pan, H., and L. Mahrt, 1987: Interaction between soil hydrology and boundary layer development. Bound.-Layer Meteor., 38 , 185–202.
Pan, H., and W. Wu, 1995: Implementing a mass flux convection parameterization package for the NCEP medium-range forecast model. NMC Office Note, 40 pp. [Available from NCEP/EMC, 5200 Auth Road, Camp Springs, MD 20764.].
Roe, G. H., 2005: Orographic precipitation. Annu. Rev. Earth Planet. Sci., 33 , 645–671.
Soong, S., and J. Kim, 1996: Simulation of a heavy precipitation event in California. Climatic Change, 32 , 55–77.
Takacs, L., 1985: A two-step scheme for the advection equation with minimized dissipation and dispersion error. Mon. Wea. Rev., 113 , 1050–1065.
Map of GPS networks in coastal Southern California as of November 2006, from the Web site of the Scripps Orbit and Permanent Array Center, University of California, San Diego. SCIGN receivers are shown by dark squares, and those of the California Real-Time Network (CRTN) are shown by gray diamonds; additional networks are indicated by other symbols. Lat–lon grid lines are shown at 2° intervals, with center of map (marked by plus sign inside square) located at 33.8°N, 118.8°W.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
Map of GPS networks in coastal Southern California as of November 2006, from the Web site of the Scripps Orbit and Permanent Array Center, University of California, San Diego. SCIGN receivers are shown by dark squares, and those of the California Real-Time Network (CRTN) are shown by gray diamonds; additional networks are indicated by other symbols. Lat–lon grid lines are shown at 2° intervals, with center of map (marked by plus sign inside square) located at 33.8°N, 118.8°W.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
Map of GPS networks in coastal Southern California as of November 2006, from the Web site of the Scripps Orbit and Permanent Array Center, University of California, San Diego. SCIGN receivers are shown by dark squares, and those of the California Real-Time Network (CRTN) are shown by gray diamonds; additional networks are indicated by other symbols. Lat–lon grid lines are shown at 2° intervals, with center of map (marked by plus sign inside square) located at 33.8°N, 118.8°W.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
PWV time series (solid lines) from near-coastal GPS receivers at (left) Harvest Platform and (right) Long Beach are compared with Eta initial fields (open circles) for 1–8 Feb 1998. Units are equivalent to millimeters.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
PWV time series (solid lines) from near-coastal GPS receivers at (left) Harvest Platform and (right) Long Beach are compared with Eta initial fields (open circles) for 1–8 Feb 1998. Units are equivalent to millimeters.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
PWV time series (solid lines) from near-coastal GPS receivers at (left) Harvest Platform and (right) Long Beach are compared with Eta initial fields (open circles) for 1–8 Feb 1998. Units are equivalent to millimeters.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
Distribution of PWV (mm) as determined from the Eta analysis for the experimental domain at 0000 UTC 2 Feb 1998. Areas with PWV greater than 20 mm are shaded; dotted line signifies lon 120°W.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
Distribution of PWV (mm) as determined from the Eta analysis for the experimental domain at 0000 UTC 2 Feb 1998. Areas with PWV greater than 20 mm are shaded; dotted line signifies lon 120°W.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
Distribution of PWV (mm) as determined from the Eta analysis for the experimental domain at 0000 UTC 2 Feb 1998. Areas with PWV greater than 20 mm are shaded; dotted line signifies lon 120°W.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
(a) Values of precipitable water vapor (mm) retrieved from the SCIGN array at 0000 UTC 2 Feb 1998. Data from San Nicolas Island receiver are particularly useful for SB–VC QPF during southerly flow events; location of La Conchita landslide (10 Jan 2005) is also indicated. (b) Vertical profiles of water vapor mixing ratios at Point Conception, from the Eta analysis (left curve) and as corrected by the GPS analysis (right curve).
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
(a) Values of precipitable water vapor (mm) retrieved from the SCIGN array at 0000 UTC 2 Feb 1998. Data from San Nicolas Island receiver are particularly useful for SB–VC QPF during southerly flow events; location of La Conchita landslide (10 Jan 2005) is also indicated. (b) Vertical profiles of water vapor mixing ratios at Point Conception, from the Eta analysis (left curve) and as corrected by the GPS analysis (right curve).
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
(a) Values of precipitable water vapor (mm) retrieved from the SCIGN array at 0000 UTC 2 Feb 1998. Data from San Nicolas Island receiver are particularly useful for SB–VC QPF during southerly flow events; location of La Conchita landslide (10 Jan 2005) is also indicated. (b) Vertical profiles of water vapor mixing ratios at Point Conception, from the Eta analysis (left curve) and as corrected by the GPS analysis (right curve).
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
(a) The PWV differences (mm) between the Eta-only and the Eta + GPS at 0000 UTC 2 Feb 1998, and the precipitation verification boxes SB and VC used in this study (dotted line identifies lon 120°W). Also shown are trajectories computed by the model for neutrally buoyant parcels released at 0000 UTC 2 Feb 1998 at an altitude of 300 m above sea level, with tick marks every 30 min. (b) Same as in (a), but instead for parcels released at an altitude of 700 m.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
(a) The PWV differences (mm) between the Eta-only and the Eta + GPS at 0000 UTC 2 Feb 1998, and the precipitation verification boxes SB and VC used in this study (dotted line identifies lon 120°W). Also shown are trajectories computed by the model for neutrally buoyant parcels released at 0000 UTC 2 Feb 1998 at an altitude of 300 m above sea level, with tick marks every 30 min. (b) Same as in (a), but instead for parcels released at an altitude of 700 m.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
(a) The PWV differences (mm) between the Eta-only and the Eta + GPS at 0000 UTC 2 Feb 1998, and the precipitation verification boxes SB and VC used in this study (dotted line identifies lon 120°W). Also shown are trajectories computed by the model for neutrally buoyant parcels released at 0000 UTC 2 Feb 1998 at an altitude of 300 m above sea level, with tick marks every 30 min. (b) Same as in (a), but instead for parcels released at an altitude of 700 m.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
Change (mm) in the 6-h QPF initialized at 0000 UTC 2 Feb 1998, due to the addition of GPS PWV data. Areas with QPF increase larger than 3 mm are shaded; dotted line identifies lon 120°W.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
Change (mm) in the 6-h QPF initialized at 0000 UTC 2 Feb 1998, due to the addition of GPS PWV data. Areas with QPF increase larger than 3 mm are shaded; dotted line identifies lon 120°W.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
Change (mm) in the 6-h QPF initialized at 0000 UTC 2 Feb 1998, due to the addition of GPS PWV data. Areas with QPF increase larger than 3 mm are shaded; dotted line identifies lon 120°W.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
(a) Cumulative observed hourly rainfall for the Santa Barbara region (Obs) compared with GPS- and Eta-only QPFs initialized at 0000 UTC 2 Feb 1998. (b) Same as in (a), but for the Ventura County region.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
(a) Cumulative observed hourly rainfall for the Santa Barbara region (Obs) compared with GPS- and Eta-only QPFs initialized at 0000 UTC 2 Feb 1998. (b) Same as in (a), but for the Ventura County region.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
(a) Cumulative observed hourly rainfall for the Santa Barbara region (Obs) compared with GPS- and Eta-only QPFs initialized at 0000 UTC 2 Feb 1998. (b) Same as in (a), but for the Ventura County region.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
Hourly rainfall for the La Conchita area (located on the coast between SB and VC) for the four days preceding the 10 Jan 2005 landslide.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
Hourly rainfall for the La Conchita area (located on the coast between SB and VC) for the four days preceding the 10 Jan 2005 landslide.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
Hourly rainfall for the La Conchita area (located on the coast between SB and VC) for the four days preceding the 10 Jan 2005 landslide.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
GPS improvements to the Santa Barbara (SB) and Ventura County (VC) 3- and 6-h QPFs leading up to the La Conchita landslide on 10 Jan 2005.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
GPS improvements to the Santa Barbara (SB) and Ventura County (VC) 3- and 6-h QPFs leading up to the La Conchita landslide on 10 Jan 2005.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
GPS improvements to the Santa Barbara (SB) and Ventura County (VC) 3- and 6-h QPFs leading up to the La Conchita landslide on 10 Jan 2005.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
Hourly Eta-only (long-dashed line) and Eta + GPS (solid line) QPFs for the La Conchita area, initialized at 1200 UTC 10 Jan 2005. Observed precipitation (short-dashed line, divided by a factor of 5 for ease of visualization), and time of the landslide are also shown.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
Hourly Eta-only (long-dashed line) and Eta + GPS (solid line) QPFs for the La Conchita area, initialized at 1200 UTC 10 Jan 2005. Observed precipitation (short-dashed line, divided by a factor of 5 for ease of visualization), and time of the landslide are also shown.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
Hourly Eta-only (long-dashed line) and Eta + GPS (solid line) QPFs for the La Conchita area, initialized at 1200 UTC 10 Jan 2005. Observed precipitation (short-dashed line, divided by a factor of 5 for ease of visualization), and time of the landslide are also shown.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
(a) Fractional variance of observed precipitation for the Santa Barbara region explained by Eta-only and Eta + GPS 3-h QPFs over the 47 cases, for gain factors of 1.0–3.0 multiplying the model forecasts. (b) Same as in (a), but rather for the Ventura County region. (c) Same as in (a), but for 6-h QPFs. (d) Same as in (b), but for 6-h QPFs. Note that mean values were removed from each 47-case ensemble before computing the explained variance.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
(a) Fractional variance of observed precipitation for the Santa Barbara region explained by Eta-only and Eta + GPS 3-h QPFs over the 47 cases, for gain factors of 1.0–3.0 multiplying the model forecasts. (b) Same as in (a), but rather for the Ventura County region. (c) Same as in (a), but for 6-h QPFs. (d) Same as in (b), but for 6-h QPFs. Note that mean values were removed from each 47-case ensemble before computing the explained variance.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
(a) Fractional variance of observed precipitation for the Santa Barbara region explained by Eta-only and Eta + GPS 3-h QPFs over the 47 cases, for gain factors of 1.0–3.0 multiplying the model forecasts. (b) Same as in (a), but rather for the Ventura County region. (c) Same as in (a), but for 6-h QPFs. (d) Same as in (b), but for 6-h QPFs. Note that mean values were removed from each 47-case ensemble before computing the explained variance.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
(a) Histogram of improvements provided by the 94 GPS-initialized forecasts (relative to the Eta-initialized forecasts) for the combined SB and VC regions (negative improvement signifies greater absolute error for the GPS QPF). (b) Same as in (a), but for the 52 cases with 300-m winds from the southern quadrant (SW–SE). Note difference in vertical scales between the two panels; mean values of the improvements are given in last four rows of Table 3.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
(a) Histogram of improvements provided by the 94 GPS-initialized forecasts (relative to the Eta-initialized forecasts) for the combined SB and VC regions (negative improvement signifies greater absolute error for the GPS QPF). (b) Same as in (a), but for the 52 cases with 300-m winds from the southern quadrant (SW–SE). Note difference in vertical scales between the two panels; mean values of the improvements are given in last four rows of Table 3.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
(a) Histogram of improvements provided by the 94 GPS-initialized forecasts (relative to the Eta-initialized forecasts) for the combined SB and VC regions (negative improvement signifies greater absolute error for the GPS QPF). (b) Same as in (a), but for the 52 cases with 300-m winds from the southern quadrant (SW–SE). Note difference in vertical scales between the two panels; mean values of the improvements are given in last four rows of Table 3.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
Normalized improvement in average QPF skill (see text) due to the addition of GPS PWV data derived from the SCIGN array, for forecast durations of 1–6 h. Results are shown for the SB and VC regions, as well as for an ensemble combining both regions and a combined subensemble characterized by low-level (300 m) winds from the southerly (SW–SE) quadrant.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
Normalized improvement in average QPF skill (see text) due to the addition of GPS PWV data derived from the SCIGN array, for forecast durations of 1–6 h. Results are shown for the SB and VC regions, as well as for an ensemble combining both regions and a combined subensemble characterized by low-level (300 m) winds from the southerly (SW–SE) quadrant.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
Normalized improvement in average QPF skill (see text) due to the addition of GPS PWV data derived from the SCIGN array, for forecast durations of 1–6 h. Results are shown for the SB and VC regions, as well as for an ensemble combining both regions and a combined subensemble characterized by low-level (300 m) winds from the southerly (SW–SE) quadrant.
Citation: Journal of Applied Meteorology and Climatology 46, 11; 10.1175/2007JAMC1502.1
Time periods of the 47 winter storm cases examined in this study for the SB and VC regions.
Average 3- and 6-h QPFs (mm) for SB and VC regions, along with corresponding ALERT rain gauge–measured precipitation, for the 47 winter storm cases. Percentage values of the QPFs in terms of observed rainfall, as well as fractional QPF change due to incorporation of GPS PWV data, are also shown.
GPS improvements to the model QPFs. Values are shown for the 47-case regional SB and VC ensembles, as well as for the combined 94-case ensembles and for the 52-case subensembles characterized by low-level winds from the southern (SW–SE) quadrant.
