FSL Forum, Nita Fullerton Editor
LINK: Intro PageLINK: Dr. Adrian MarroquinLINK: Cycled Snow VariablesLINK: GFE at NWSLINK: Defining Observed FieldsLINK: WRF Model in NWSLINK: Value of Profiler DataLINK: Evaluating ForecastsLINK: BriefsLINK: Recent Publications  
By Tatiana Smirnova, Stanley Benjamin, John Brown, and Dongsoo Kim



Accurate estimation of snow water equivalent over the U.S. and adjacent areas is critical for subsequent seasonal and short-range atmospheric and hydrological model forecasts. It is also essential for a variety of important water planning activities in the western U.S., including agriculture, recreation, and public use in cities. It is difficult to estimate precipitation in much of the western U.S. due to complex terrain. For example, quantitative precipitation estimates (QPEs) based only on observations are often deficient in the cold season, especially for orographic precipitation. Recent approaches toward development of land data assimilation systems and precipitation assimilation used in the National Centers for Environmental Prediction (NCEP) Regional Reanalysis are largely focused on improvements in the specification of the land-surface state for the warm season, but are not generally applicable to the problem of snow water equivalent initialization.

The best QPEs now available – those used to drive the current NOAA/NASA Land Data Assimilation System (LDAS) – are taken from the NOAA Climate Prediction Center's 24-hour gauge precipitation analysis. These QPEs are allocated into hourly amounts using the NCEP/National Weather Service (NWS) Stage IV hourly precipitation analysis. However, this analysis has limitations, including insufficient density of rain gauges in the mountainous areas, day-to-day variations of the reporting gauges, errors in 24-hour time variations, inaccuracies of gauge observations for frozen precipitation, and difficulties in assigning the precipitation phase.

FSL has developed a four-dimensional Coupled Data Assimilation System (CDAS) using a forward, full-physics model in which the precipitation and clouds are an optimized combination of observed and forecast fields to reduce the negative effects from the limitations of the precipitation analysis mentioned above. Based on the Rapid Update Cycle (RUC) model assimilation system, CDAS updates the precipitation and cloud fields hourly using GOES cloud-top pressure, NEXRAD radar reflectivity, lightning data, and GPS precipitable water. The RUC CDAS is designed specifically to provide improved quantitative precipitation estimates for orographic precipitation in the cold season, leading to more accurate cycling of the snow state over the U.S.

Description of the RUC CDAS Forecast Model

The RUC CDAS is a coupled land-surface/atmospheric model with an hourly assimilation cycle including radar reflectivity, satellite, and other remotely sensed data to update the three-dimensional hydrometeor fields evolving through explicit mixed-phase cloud microphysics in the RUC model. The RUC is the only NCEP operational model that currently provides explicit forecasts of precipitation in liquid and solid phases. Model precipitation has consistent spatial variability from day to day, and could therefore be used to mitigate the effect of missing stations. It is also likely to provide better orographic precipitation, especially if constrained by satellite, radar, and even gauges in the CDAS optimization. Because of the 1-hour updating used in the RUC, the model is also constrained by the hourly input of surface observations to have fairly accurate short-term forecasts of low-level temperatures. The explicit microphysics and representation of the near-surface thermal structure of the atmosphere in the RUC may be expected to provide better information on precipitation type than that available from an estimate of surface temperature alone.

Precipitation Physics

Explicit Mixed-Phase Cloud Microphysics – The RUC CDAS uses the improved version (Thomspon et al. 2003) of the level 4 mixed-phase bulk microphysics scheme of Reisner et al. (1998). This scheme was originally developed for MM5, and has been used operationally in the RUC for the past 5 years (Brown et al. 2000). The mixing ratios of 5 hydrometeor species are explicitly predicted: cloud water, rain water, cloud ice, snow and graupel (the latter formed by riming of ice or snow, or by freezing of rain drops, in which circumstance it might be regarded as ice pellets or sleet); the number concentration of cloud ice particles is also predicted. The explicit prediction of these hydrometeors allows direct feedback between simulated clouds and long- and short-wave radiation. The RUC/MM5 explicit microphysics also allows an explicit forecast of snow and snow water equivalent, rather than a diagnostic for precipitation phase type based on temperature.

The original version of the scheme used in the RUC tended to strongly overpredict graupel under certain conditions, leading to unrealistic distributions of surface-precipitation type. The new version, now operational in the RUC, exhibits much improved predictions of supercooled liquid water, as well as of precipitation type at ground level. Continued enhancements to the RUC microphysics scheme are expected, with a focus on ice nucleation and explicit prediction of freezing drizzle in weakly forced synoptic situations.

Ensemble Cumulus Parameterization – A new convective parameterization of Grell and Devenyi (2002) is now used in the RUC. The original scheme was first expanded to include lateral entrainment and detrainment, including detrainment of cloud water and ice to the microphysics scheme mentioned above. In addition, the scheme draws on uncertainties in convective parameterizations by allowing an ensemble of various closure and feedback assumptions (related to how the explicitly predicted flow modifies the parameterized convection, which in turn modifies the environment) to be used every time the parameterization is called.

The four main groups of closures that are used in the RUC application are based on removal of convective available potential energy (CAPE), destabilization effects, moisture convergence, and low-level vertical velocity. These four groups are then perturbed by 27 sensitive parameters related to feedback as well as strength of convection, which give a total of 108 ensemble members that contribute to the convective scheme. Output from the parameterization may be the ensemble mean, the most probable value, a probability density function, as well as other statistical values. Currently, only the ensemble mean is fed back to the dynamic model. The application of the Grell/Devenyi convective scheme in the RUC model also includes a removal of the negative buoyancy capping constraint at the initial time of each model forecast in areas where the GOES sounder's effective cloud amount indicates that convection may be present. This technique can aid the initiation of modeled convection at grid points where positive CAPE is observed, although it cannot create positive CAPE. In addition, an upstream dependence is introduced through relaxation of static stability (convective inhibition) constraints at adjacent downstream points based on 0–5 km mean wind, and through allowance of the downdraft mass flux at the previous convective timestep to force convection at the downstream location.

Land-Surface Model in the RUC

The sophistication of the RUC/MAPS land-surface model (LSM, Smirnova et al. 2000, 1998) has also grown in the past few years, and is now used in the operational RUC at NCEP, in experimental real-time RUC runs at FSL, and in regional climate versions of the RUC and MM5 models. The RUC LSM has also made a strong performance in the Program for Intercomparison of Land-surface Process Models (PILPS, Schlosser et al. 1999) and in the intercomparison of the snow models (SNOWMIP, Etchevers et al. 2003). The RUC LSM contains multilevel soil and snow models, and treatment of vegetation (Figure 1), all operating on the same horizontal grid as the atmospheric model. Heat and moisture transfer equations are solved at six levels for each soil column together with the energy and moisture budget equations for the ground surface, and an implicit scheme is used for the computation of the surface fluxes. The energy and moisture budgets are applied to a thin layer spanning the ground surface and including both the soil and the atmosphere with corresponding heat capacities and densities. The RUC frozen soil parameterization considers latent heat of phase changes in soil by applying an apparent heat capacity, augmented to account for phase changes inside the soil, to the heat transfer equation in frozen soil in place of the volumetric heat capacity for unfrozen soil. The effect of ice in soil on water transport is also considered in formulating the hydraulic and diffusional conductivities.

Figure 1. RUC/MAPS land-surface model.

Figure 1. RUC/MAPS land-surface model.


Accumulation of precipitation at the surface, as well as its partitioning between liquid and solid phases, is provided by the mixed-phase cloud microphysics routine. In the RUC, the convective parameterization also contributes to the snow accumulation if the surface air temperature is at or below 0°C. With or without snow cover, surface runoff occurs if the rate at which liquid phase becomes available for infiltration at the ground surface exceeds the maximum infiltration rate. The solid phase in the form of snow or graupel (treated identically by the LSM) is accumulated on the ground/snow surface to subsequently affect soil hydrology and thermodynamics of the low atmosphere.

In the most recent version of the LSM implemented in the RUC, a number of improvements were achieved in the treatment of snow cover and frozen soil physics. These improvements include the allowance of the evolution of snow density as a function of snow age and depth, the potential for refreezing of melted water inside the snowpack, and simple representation of patchy snow through reduction of the albedo when the snow depth is small. If the snow layer is thinner than a 2-cm threshold, it is combined with the top soil layer to permit a more accurate solution of the energy budget. This strategy gives improved prediction of nighttime surface temperatures under clear conditions and melting of shallow snow cover. Another feature of the RUC LSM is an improved algorithm for frozen soil physics for spring thaw conditions.

These changes were tested off-line in a one-dimensional setting with the dataset from Valdai, Russia, and showed positive impact on the model performance. The evolution of snow density provides a more realistic representation of processes in snow, especially when fresh snow is falling onto the bare soil or an existing snowpack, and improves simulation of the snow-melting season (Figure 2). The effects of patchy snow cover were tested in the experimental version of RUC and improved prediction of the nighttime surface temperatures under clear conditions as well as melting of shallow snow cover.


Figure 2. Results from one-dimensional simulations with the RUC/MAPS land-surface model for winter 1980 – 1981 for Valdai, Russia (PILPS 2D experiment, 18-year simulation). Results show improvement from allowing variability of snow density and adding the Johansen formulation for thermal conductivity. (a, above) Total runoff (millimeters) from top 1 meter of soil; (b, below) snow water equivalent (millimeters).


More accurate predictions of the surface temperature have positive effects on the verification of 2-m temperature (Figure 3). We will also investigate adding improvements to the vegetation treatment such as interactive vegetation in the RUC land-surface model to improve its capability for regional climate prediction and simulation.


Figure 3. Surface temperature biases (oC) from 3-hour forecasts for stations with the snow depth less than 10 centimeters averaged for the period 4 – 14 February 2001.

In applications of the RUC LSM in current and previous versions of the RUC, volumetric soil moisture and soil temperature are cycled at the six soil model levels, as well as canopy water, snow depth, and snow temperature. Cycling of the snow temperature of the second layer (where needed) is also performed. The RUC continues to be unique among operational models in its specification of snow cover and snow water content through cycling (Smirnova et al. 2000). The 2-layer snow model in the RUC improves the evolution of these fields, especially in springtime, more accurately depicting the snow melting season and spring spike in total runoff.

Evaluation of the RUC CDAS

The RUC CDAS has a 20-km resolution, with high-resolution fixed surface fields from the USGS land-use and STATSGO (State Soil Geographic database) soil types. This system has run in real time since April 2002. The RUC/MAPS system made a transition to a three-dimensional variational (3DVAR) analysis (Devenyi and Benjamin 2003) in spring 2003 replacing the previous optimal interpolation atmospheric analysis scheme. The primary reason for this change is because of the flexibility and rigor of the variational approach in assimilating data not directly forecast by the model (such as satellite radiances and radar radial winds). The RUC CDAS currently assimilates hourly data wind profiler data from the NOAA Profiler Network and the 915-MHz Boundary Layer Network, commercial aircraft (growing rapidly in volume over the U.S. and worldwide), surface stations and buoys, available radiosonde data, and GOES satellite-estimated cloud drift winds, precipitable water measurements, and cloud-top pressure. Assimilation of GPS integrated precipitable water measurements has undergone testing for three years with the RUC and also has been implemented operationally. The most critical recent improvements in RUC assimilation were done to the cloud and moisture analysis through the assimilation of satellite and radar data.

The RUC CDAS provides refinements refinements to the 0–1 hour precipitation forecasts that drive a land-surface climate in the model by accounting for errors in both observations and model precipitation forecasts. As a result, cycled soil moisture and snow water equivalent are improved compared to the operational RUC without assimilation of the radar, lightning, and GPS data. Most improvements occur from more accurate placement of predicted precipitation.

The impact of using the radar assimilation technique to modify 3D hydrometeor fields from the national mosaic 2-km resolution maximum reflectivity data has been regularly monitored and evaluated. The operational RUC20 – without radar reflectivity assimilation – is used as the control experiment.

The NCEP Stage II hourly QPE is used to verify the 3-hour accumulated precipitation, and quality controlled Stage IV is used for the verification of 24-hour accumulations. The Stage IV precipitation data, derived from NEXRAD reflectivity and gauge observations, are at 4-km resolution and include quality control. The original 4-km resolution Stage IV precipitation data are remapped to the RUC grid by taking the maximum value in the grid box to represent the grid point. The verification of accumulated forecast precipitation from eight 3-hour forecasts in RUC assimilation cycles over a 24-hour period is performed daily (see the example in Figure 4. )


Figure 4. A 24-hour accumulation of precipitation for the period ending 1200 UTC 7 April 2003 from (a, top left) control run, without radar reflectivity assimilation, (b, middle left) from parallel run with radar reflectivity, and (c, bottom left) Stage IV precipitation amounts (sum of four 6-hour totals). Forecast amounts are for 8 consecutive 3-hour forecasts from RUC cycles. (d, e, and f, right panel) are spatial correlation functions corresponding to a), b), and c), and the maximum correlation is also shown. See text for more explanation.

A spatial correlation field was computed as a measure of precipitation verification. The spatial cross-correlation is a function of x-y displacement between two fields, QPF and QPE within a predetermined evaluation window (60 x 60 grid points on a 20-km grid, Figure 4d–f). The distance of maximum correlation to the center (zero displacement) is a measure of QPF phase error, and the maximum value of correlation coefficient provides an approximate measure of forecast accuracy modulated by spatial variability of rainfall amount. The shape of the contours gives information on the directional dependency of precipitation forecast accuracy.

The two contour fields were compared with the spatial autocorrelation field, which is computed from the QPE against itself. The preferred orientation of precipitation isopleths during this period is evident, with strong anisotropy oriented from west-southwest to north-northeast. The spatial patterns also depend on the duration of accumulation. As an overall assessment, better QPF should result in a QPF-QPE correlation pattern similar to that of the spatial auto correlation. Figure 4 shows that the maximum value of cross correlation coefficient of the parallel run (with radar reflectivity assimilation) is 0.67, better than 0.58 for the control run (without radar reflectivity assimilation), indicating that the QPF error in the parallel run is reduced from that of the control run. The contour lines of the parallel run result are better defined, suggesting that its spatial scales and directional orientations are more accurate than those of the control run in this case.

Snow State Cycling in the RUC CDAS

The RUC Control and RUC CDAS were run in parallel during the first cold season of winter 2002-2003. The advantages of of the RUC CDAS could be monitored in the evolution of the snow depth field driven by the 1-hour precipitation forecast. Although the atmospheric forcing from the RUC 1-hour cycle often corrects misplaced snow precipitation by providing the energy for snow melting, still the snow field is a very good indicator of the improvements in the precipitation forecasts in the RUC CDAS.

The example in Figure 5 (a-c) illustrates the comparison of snow depths from the RUC Control and RUC CDAS cycles against the NOHRSC (National Operational Hydrologic Remote Sensing Center) NSA (National Snow Analysis). The NOHRSC product combines the snow model assimilation with all available snow observations and provides one of the most reliable datasets of this variable. Although RUC CDAS improves the cycled snow state, at the same time certain deficiencies still exist in the amounts of the precipitation in the RUC CDAS, causing most often underestimation of cycled snow depth. This also has a delayed effect on the soil moisture climate and surface physics in the warm season.


Figure 5. Snow depth from the RUC Control (a, above left) and RUC CDAS (b, above right) verified against the NOHRSC snow analysis (c, right) valid on 30 January 2003.


Further improvements of the cycled snow depth could be achieved by updating the snow state fields from existing observations. This approach has been implemented at the National Operational Hydrologic Remote Sensing Center in their National Snow Analysis. Snow data used to update the model include observations from the NOHRSC's Airborne Snow Survey Program, NWS and FAA field offices, NWS Cooperative Observers, the National Resources Conservation Service (NRCS) Snow Water Equivalent Information (SNOTEL) and snow course networks, the California Department of Water Resources snow pillow networks, and snow cover observations from NOAA's GOES and AVHRR satellites.

The first step in this direction is to compare RUC CDAS snow state variables to the NOHRSC NSA and identify the areas with the largest deficiencies. Then the technique should be developed and applied to make corrections of RUC CDAS snow variables for these areas. Improvements of the snow climate in the RUC CDAS will also be beneficial for the NOHRSC NSA, because that snow model is driven by the RUC precipitation and atmospheric forcing. An example of time series products from the NSA showing the verification of the RUC precipitation forcing as well as NSA snow depth verification for the March 2003 snow storm in Boulder, Colorado, is presented on Figure 6 (a,b). In this particular case, the RUC model was able to provide sufficiently accurate forcing for the NOHRSC snow model, and corrections of the snow analysis from observations were not needed. Similar verification of the RUC precipitation and atmospheric forcing is performed regularly at the NOHRSC for different stations, and it demonstrates that in some cases the improvements to the RUC precipitation forcing are necessary. We will continue to make more detailed comparisons between the RUC CDAS snow state variables and the NOHRSC NSA, and these results will be presented in future articles.


Figure 6. The precipitation forcing from RUC 1-hour forecasts (a, left top panel) for March 2003 snow storm compared to the observed precipitation (a, left bottom panel), and the National Snow Analysis snow depth (b) comparison to observations for Boulder, Colorado.


Note: A complete list of references and more information on this and related topics are available at the main FSL Website, by clicking on "Publications" and "Research Articles."

(Tanya Smirnova is a researcher (under CIRES contract) in the Regional Analysis and Prediction Branch, headed by Dr. Stan Benjamin. She can be reached by email at, or by phone at 303-497-6253.)