1. Introduction
The design of hydrometric networks is a classical problem in hydrometeorology. Central to this issue is that benefits produced by the possession of data should never be less than the data collection costs themselves (Rodríguez-Iturbe and Mejía 1974). One can approach the problem by either eliminating redundant stations to reduce costs or augmenting a network to increase benefits, and studies have examined these strategies using a variety of methods for a range of data collection objectives (see Mishra and Coulibaly 2009).
With respect to surface water networks, a number of network design studies have been motivated by the need for more accurate prediction of streamflow, such as for flood hazard management (Tsintikidis et al. 2002; Volkmann et al. 2010) or water supply (Peck and Schaake 1990). The methodology used by these studies is necessarily dictated by the prediction model for which the network data are intended. Because many of these prediction models are either physically based or lumped conceptual runoff simulation models, an implicit assumption is that more accurate estimates of mean areal precipitation result in more accurate streamflow predictions, and often interpolation methods such as kriging are applied to historical observations for the network design approach (Pardo-Igúzquiza 1998; Tsintikidis et al. 2002; Volkmann et al. 2010). These methods, however, may be less appropriate for statistical prediction models based on point observations of hydrologic variables.
One such prediction model is that of the Natural Resources Conservation Service (NRCS) of the U.S. Department of Agriculture. Since 1935, a principal responsibility of NRCS has been the publication of water supply forecasts in the American West (Helms et al. 2008). End users of the forecasts serve a broad array of objectives ranging from irrigated agriculture, flood control, and municipal water supply to endangered species protection, power generation, and recreation (Pagano et al. 2009). Traditionally, NRCS has coordinated with the National Weather Service (NWS) to publish forecasts at the beginning of each month from January through June for several hundred locations throughout the region.
To generate forecasts, NRCS relies on a multivariate principal components regression (PCR) methodology based primarily on point observations of initial hydrologic conditions (IHCs; Garen 1992). IHCs are observed by an extensive network of approximately 1200 snow courses and 850 Snowpack Telemetry (SNOTEL) sites scattered throughout the western United States [D. Garen et al., NRCS National Water and Climate Center (NWCC), 2012, personal communication]. Over the past 10–15 years, about 40% of SNOTEL stations have been retrofitted with sensors of soil moisture and several other environmental parameters (so-called “enhanced” SNOTEL stations). Data from the NRCS Soil Climate Analysis Network (SCAN; Schaefer et al. 2007) are also available, although most records are still too short for use in statistical forecasts.
New SNOTEL installations have averaged 13 per year since 437 stations were installed in 1979 and 1980, with the frequency of new installations varying highly from state to state. Equipment costs for a standard installation total $25 000, a sum that is quickly outpaced by annual operation and maintenance costs, which are amplified by remote site locations and often challenging to finance (X. XXXXXX, NRCS NWCC, 2012, personal communication). Site selection for new SNOTEL stations is generally an ad hoc process that may be influenced by the offer of funding cooperation from a user group, land availability, and a qualitative perception of monitoring network needs. For years, the guiding philosophy was simply to situate new stations at existing snow courses—themselves likely sited based primarily on local hydrologic knowledge and site accessibility—providing continuous monitoring at locations that otherwise would have monthly or less frequent reports and enabling intercomparison between SNOTEL and snow course observations. A geographic information system (GIS) tool was recently developed to identify monitoring gaps (Perkins et al. 2010), but it does not employ quantitative metrics that evaluate impacts of new stations on streamflow forecast accuracy. However, budgetary limitations argue that network augmentation to support hydrologic prediction should prioritize drainage basins for which skillful forecasts are most critical and local needs are least satisfied and identify sites with the best potential to offer skill improvements.
This paper presents a hydrometric network design approach toward the objective of enhancing statistical prediction models. The specific focus of the paper is the development of a forecast skill-oriented technique for informing NRCS SNOTEL network expansion decisions. We employ a hybrid dynamical–statistical approach that combines the dimension-reducing power of the NRCS PCR methodology with the spatially distributed nature of a physically based, macroscale hydrology model. Principal components analysis is a long-standing network design technique in its own right (Fiering 1965; Morin et al. 1979) and, as described previously, physically based hydrologic models have been used extensively in these types of studies as well. The innovations presented herein are in the combination of these techniques for network design and the use of simulated data as surrogates for point observations in the prediction model.
2. Study areas
A collection of study basins was selected to represent the diversity of physiographic, climatic, and existing operational network conditions across the western United States. Selection criteria stipulated that watersheds be headwater basins, since forecasts for points farther downstream are typically based on routed relationships with upstream forecasts rather than point observations of IHCs. An introductory meeting was held at the NRCS NWCC offices in Portland, Oregon, and, given only this requirement, each of the five NWCC hydrologists were asked to identify up to five watersheds of interest from their respective forecast regions. In total, 24 study basins were selected from six water resource regions (Table 1), with reasons for selection tending to vary among hydrologists and basins. Some, such as DTTM8 and HLWM8 in the Missouri water resource region, were chosen because of highly variable climatologies, rendering forecasting for these basins more difficult. Others are critical forecast points for larger river systems such as the Klamath (CHSO3), Bighorn (CROW4), and North Platte (NGTC2). Those in the Colorado and Rio Grande were selected as a general cross section of the major tributaries across their respective regions. Still others, such as MONO3 and PRTI1 in the Pacific Northwest, were chosen because of sparse existing networks but numerous options for new installations.
Identification data for the 24 basins in the study; R stands for river and Res stands for reservoir.
Table 2 presents various physiographic and climatic statistics for each of the basins in the study. Study basin locations are shown in Fig. 1, which also includes a map of the National Wilderness Preservation System, for which observation equipment is restricted (Landres et al. 2003). As shown, several of the study basins occupy land within this system. Thus, an additional motivation in this study is to determine whether parts of these wilderness areas are important for seasonal streamflow prediction. With about 90% of its area falling within this system, HHWM8 is a classic example of this scenario, but others (e.g., EGLC2, GBRW4, and NVRN5) contain wilderness areas as well, primarily at higher elevations near basin boundaries.
Summary statistics for the 24 basins in the study. Mean values were calculated over the calibration period. The annual runoff ratio is defined as the ratio of annual runoff to annual precipitation.
The 24 basins in the study. Water resource regions are delineated by dashed lines, and federally protected wilderness areas are shown in dark gray.
Citation: Journal of Hydrometeorology 14, 5; 10.1175/JHM-D-12-0136.1
3. Data and methods
We formulated our approach to specifically address the question, “Given an existing hydrometric network for a particular watershed, where is the next best location to install a SNOTEL station?” The elements of this approach are detailed in the sections that follow.
a. Forecast methodology
1) NRCS procedure
The statistical forecasting approach of NRCS treats each point (station) observation as an independent predictor in forecasts of seasonal streamflow. Principal components are used to circumvent the well-known problem of predictor collinearity in multivariate regression. Forecast models are developed by restructuring predictor variables into principal components, arranging principal components in order of decreasing eigenvalue (explained variance), developing an equation that sequentially retains only those principal components deemed significant via a t test, and inverting the transformation so that coefficients are expressed in terms of the original predictor variables (Garen 1992; Garen and Pagano 2007). In addition to snow water equivalent (SWE) and water year-to-date precipitation from NRCS snow courses and SNOTEL stations, other predictors such as NWS Cooperative Observer precipitation observations, U.S. Geological Survey (USGS) streamflow observations, and teleconnection indices are occasionally used.
The NRCS procedure employs an iterative search routine that optimizes variable combinations by developing all possible equations resulting from an increasing number of predictors. With each additional variable, the jackknifed standard error (JSE) is used to order the equations, and the top 30 equations are identified. When the top 30 equations cease to evolve, a final equation is selected by balancing the objectives of a low JSE and month-to-month variable consistency (Garen 1992; Garen and Pagano 2007). The resulting median predictions are combined with the JSE to support probabilistic water supply forecasts that are released to the public in the form of five quantiles (10, 30, 50, 70, and 90).
For the present study, current operational forecast equations and historical naturalized streamflow observations were obtained directly from NRCS. Table 2 lists the forecast target periods for each of the basins; for those in which forecasts are issued for multiple target periods, the April–July period was arbitrarily selected for the study. Historical predictor data were downloaded from www.wcc.nrcs.usda.gov/reportGenerator. Calibration periods for the forecast models described herein were set to water years 1981–2010 unless insufficient predictor or predictand data were available, in which case the calibration period was adjusted as listed in Table 2. Only median forecasts were considered for purposes of analysis.
2) Hybrid approach
The hybrid framework of Rosenberg et al. (2011) exploits the distributed nature of macroscale hydrologic models to expand predictor sets for statistical forecasting applications. As developed for the pilot study of California's Central Valley drainage, the method uses the Variable Infiltration Capacity (VIC) macroscale hydrology model (Liang et al. 1994) to simulate SWE at a 
In this study, we employed the
b. Exploratory assessment
1) Screening of variables
We examined the predictability afforded by SNOTEL-observed variables to determine those to include in the network design. Variables known to influence the water balance, and consequently seasonal streamflow, were evaluated. Candidate variables included the traditional predictors of SWE and water year-to-date accumulated precipitation and also two additional variables, soil moisture and air temperature. Soil moisture has long been recognized for its potential to improve seasonal streamflow forecasts (Boardman 1936; Clyde 1940) and has also been the subject of some recent work on SNOTEL data (Lea and Harms 2011), but it has not been directly incorporated by NRCS in an operational statistical framework. Air temperature influences the cold content of snow, which in turn affects the timing and possibly the efficiency of snowmelt runoff (Speers et al. 1996). Notwithstanding some short-term forecasting applications (Tangborn 1980), however, daily temperature observations, which are available from most SNOTEL stations, have rarely been examined in a statistical runoff prediction context.
The comparison of predictor skill was based entirely on VIC forcings and simulated hydrologic variables. Accumulated water year-to-date precipitation and average water year-to-date temperature were calculated at monthly intervals from daily VIC forcing data, which were interpolated from NWS Cooperative Observer station data and scaled (in the case of precipitation) to match monthly climatologies from the Parameter-Elevation Regressions on Independent Slopes Model (PRISM; Maurer et al. 2002). The VIC simulations provided soil moisture data for each of three model soil layers (termed SM1, SM2, and SM3) on the first of each month. First-of-the-month SWE data were drawn from simulations for up to five elevation bands for each grid cell, depending on the grid cell's elevation range. Basinwide data were compiled for the calibration periods indicated in Table 2 and processed through the PCR algorithm described in section 3a(1). Results corroborated the well-known predictive role of SWE, particularly for those basins at higher elevations. Soil moisture and precipitation, which (like streamflow) can be viewed as a proxy for soil moisture in statistical forecasts (Speers et al. 1996), demonstrated comparable and useful skill in most basins. Temperature, on the other hand, despite several attempts with variations of data, added little skill to forecasts and was excluded from further analysis.
2) Comparison with observations
A central premise of the study is that gridded/simulated data are suitable proxies for observed conditions. Although some studies have suggested that SNOTEL SWE values are generally unrepresentative of grid element SWE (Molotch and Bales 2005), a more meaningful measure for statistical forecasts is the correspondence of variance. Accordingly, we computed correlations between SWE and precipitation observations from existing NRCS stations in each of the 24 basins and gridded/simulated data for the nearest grid cell. Because PRISM data are partially based on NRCS SNOTEL observations, precipitation correlations were expected to be high. Computations were performed only on data relevant to operational forecasts, those occurring on the first of the month.
For soil moisture, SNOTEL stations within each of the study basins were scanned for historical observations, and soil moisture records beginning in water year 2006 or earlier were obtained (checks of SCAN stations did not result in any meeting this criteria). This ensured at least five years of data for comparison with simulations, which extended through water year 2010. In total, soil moisture observations were obtained for 59 SNOTEL stations in 17 basins. Each station had records for at least three depths—typically 2, 8, and 20 in. (~50, 200, and 500 mm)—and occasionally 4, 11, and 40 in. (~100, 280, and 1000 mm). Soil moisture correlations were computed for the simulated soil layer nearest in depth to each observation (typically SM1 for 2/4 in., SM2 for 8/11/20 in., and SM3 for 40 in.).
c. Experimental approach
1) Augmented predictor analysis
Two experiments were used to examine the potential for improvement beyond current operational forecast skill. In the first (EXP1), representing a “business as usual” scenario, gridded precipitation and simulated SWE for all grid cells and elevation bands in each basin were added to NRCS predictors in the pool of predictor candidates. In the second (EXP2), basinwide soil moisture simulations for each of the three soil layers were added to the pool of predictor candidates in EXP1. The results of these experiments estimate the limits of skill that are possible by completely sampling major water balance variables in each basin at the modeled spatial resolution of
2) Network design
The network design approach was formulated as a variation on the experiments described above. Instead of adding modeled data for all grid cells to the pool of predictor candidates at once, we iterated through the grid cells and sequentially added data from just one grid cell at a time. This ensured that improvements in forecast skill were due exclusively to predictor data from that grid cell, whereas in the augmented predictor analysis, it would have been impossible to isolate the effects of a single gridded/simulated predictor. Any improvement in JSE over that of the baseline pool of NRCS predictors was then recorded. If multiple elevation bands were present for a given grid cell, this step was repeated for each of the elevation bands, and only the largest improvement was recorded.
We then identified the modeled predictor variable(s) underlying each improvement and generated spatial images of key statistics for each of the five predictor types (SWE, water year-to-date precipitation, SM1, SM2, and SM3). Results were tallied by computing the grid cell offering the best forecast improvement each month and the grid cell offering the best average improvement over all forecast months, which was selected as the best location for a new station. This latter computation was performed in terms of millions of cubic meters, rather than percentage of mean target period streamflow, so as not to give undue weight to forecast improvements later in the snowmelt season.
d. Evaluation metrics
As a means of validating modeled snow cover data, we used imagery from the Moderate Resolution Imaging Spectroradiometer (MODIS), specifically the MOD10C2 dataset available since February 2000 (Hall et al. 2006). This dataset consists of fractional snow cover data at a 0.05° spatial resolution for 8-day periods, offering a significant reduction in the number of cloud-obscured pixels over the daily snow cover products. For each of the locations selected in the network design that were associated with a SWE predictor, we determined the nearest MODIS pixel to the respective VIC grid cell and extracted data for the 8-day period that fell nearest the first of each month (over water years 2001–10) for which simulated SWE were included in the prediction model. We then performed binary snow cover comparisons between the MOD10C2 and VIC-simulated data, assigning a value of 1 to any nonzero MODI10C2 fractional snow cover data and any nonzero simulated SWE data.
e. Operational application
Finally, we evaluated the potential impacts of a new SNOTEL station in the context of operational forecast skill. Recall that an important difference between the NRCS forecasting approach and the one adopted herein involves the NRCS practice of maintaining consistent predictors from month to month. To account for this discrepancy, we reformulated our optimized forecast models to conform to this approach, that is, forced these models to include all current NRCS predictors in addition to gridded/simulated data from the grid cell selected in the network design analysis. We employed some discretion in these models and only included the new predictor data if an improvement in skill resulted.
As an additional basis for comparison, we generated seasonal streamflow hindcasts (forecasts generated retrospectively) for study basins in the Upper Colorado water resource region. Hindcasts were produced using current (as of 2012) operational implementations of the NWS ensemble streamflow prediction (ESP) and statistical water supply (SWS) methodologies for the same periods used in the calibration of our forecast models (Table 2). ESP involves forcing NWS hydrology models with observed meteorological data to initialize (spin up) watershed states on the start day of the forecast, and then with historical sequences of meteorological data beginning on the same Julian day, to generate an ensemble of projected streamflows (Day 1985). The SWS approach (Hartman and Henkel 1994) is essentially identical to the PCR methodology of NRCS. For consistency with the rest of the analysis, the medians of these hindcasts were then selected for comparative purposes. We also obtained historical as-issued NRCS/NWS coordinated forecasts as available (R. Tama, NRCS NWCC, 2012, personal communication). Because of its utility in comparisons of this type, here we used R2 between predicted and observed streamflows as our skill metric.
4. Results
We present the results of our analyses in the order they were described above. Section 4a provides results from the correlation analysis along with skill results for the augmented predictor analysis and network design exercise. Section 4b examines the network design results in a geospatial context, while section 4c summarizes the snow cover comparisons and analysis of predictor contributions. Section 4d provides a final assessment of forecast skill from an operational perspective.
a. Correlation and skill analysis
A summary of the correlations between station observations and their gridded/simulated counterparts is shown in Fig. 2. In general, correlations for SWE and precipitation were strong, while those for soil moisture were more divergent, a result consistent with other studies (Koster et al. 2009). Thus, the correspondence of observed and simulated variables generally supported the network design assumptions, though somewhat less so for soil moisture.
Ranges of correlations between observed and gridded/simulated predictor data. The numbers of observed data records are shown at top for their respective predictor types.
Citation: Journal of Hydrometeorology 14, 5; 10.1175/JHM-D-12-0136.1
Figure 3 compares the skill of the different forecasting approaches in the augmented predictor analysis in terms of relative standard error (standard error as a percentage of the mean of the predictand). Here the predictand is streamflow for a shrinking target period, that is, only the portion of the entire target period remaining for a given forecast month. Differences between NRCS forecast skill with and without the month-to-month consistency requirement are generally small, with the consistency requirement degrading skill between near zero (HALI1, PRTI1) and ~10% (HREN2). In most basins, improvements over the baseline equations (NRCS predictors without the month-to-month consistency requirement) are fairly nominal for EXP1, with the more pronounced occurring in basins that are sparsely sampled (e.g., HLWM8) or later in the snowmelt season as in Rosenberg et al. (2011). Improvements for EXP2, however, are more substantial, particularly in basins such as DURU1 and HREN2.
Forecast skill for the augmented predictor analysis. The y axis represents standard error as a percentage of mean streamflow for a shrinking target period. Purple lines reflect operational NRCS forecast skill; gray lines reflect forecasts for which the requirement of month-to-month predictor consistency is relaxed.
Citation: Journal of Hydrometeorology 14, 5; 10.1175/JHM-D-12-0136.1
Figure 4 presents results from the network design analysis. In general, improvements resulting from the inclusion of predictor data for a single additional grid cell are proportional to those found from the inclusion of predictor data for all grid cells basinwide, a correspondence that is not surprising. Although improvements are typically possible for at least one grid cell in every forecast month, the grid cell offering the best improvement overall tends to improve forecasts either during the accumulation season or the ablation season, but rarely both. For EXP1, for example, ablation season improvements occur in BMDC2 and DTTM8, while accumulation season improvements occur in HLWM8 and HREN2. For EXP2, most of the best grid cells overall offer improvements exclusively during the accumulation season. This is a consequence of the locations of these critical grid cells, which typically differ for the two seasons as described in greater detail below.
Forecast skill for the network design analysis. The y axis is as in Fig. 3, with various labeling schemes as indicated. Note that the lines labeled “NRCS predictors” and “all grid cells” (blue for EXP1 and brown for EXP2) match those in Fig. 3. For some plots, lines may be obscured by others with similar values.
Citation: Journal of Hydrometeorology 14, 5; 10.1175/JHM-D-12-0136.1
b. Geospatial analysis
Analysis of the spatial patterns associated with the improvements provides insight into the hydrologic mechanisms from which they arise. For a representative set of study basins, Fig. 5 shows the locations of grid cells that resulted in forecast improvements for a given month, including the best grid cells that month and over all months. Results for June forecasts using EXP1 predictors at BMDC2 are typical of the late-season improvements that occurred in some of the basins. Comparison of this map with its corresponding plot in Fig. 4 indicates that the best grid cell for that month provides an improvement of 5%–10% in relative standard error; the best grid cell overall is situated nearby. Both are located at relatively high elevations, though lower than the basin's peak elevation. Inspection of the top two rows in Fig. 6 reveals information about the predictor variables underlying these improvements, the most dominant of which is SWE. The leftmost spatial image depicts average 1 June SWE for those grid cells that were considered predictor candidates [see section 3a(2)]. The climatological value for both best overall and month-optimized locations is less than ~200 mm, with the latter location falling on the fringe of the zone with any measurable SWE that time of year. The right two plots in this row reveal that patterns for correlation (with the predictand) tend toward the inverse of those for coefficient of variation (CV), a result noted for other basins as well. The selected locations fall in areas of relatively high correlation, but also somewhat higher CV than most of the other grid cells considered as predictor candidates, indicating that these locations add some measure of SWE variability to the predictor pool.
Selected spatial results from the network design analysis. The sizes of the black circles are proportional to the improvements in forecast skill offered by their respective grid cells, and the green and red circles correspond to the points on the green and red lines at the given forecast months in Fig. 4.
Citation: Journal of Hydrometeorology 14, 5; 10.1175/JHM-D-12-0136.1
Spatial statistics for the dominant predictor types corresponding to the forecast improvements in Fig. 5. Grid cells for which the specified predictor type was selected for the forecast model are indicated in the leftmost plots by black dots, with the green and red dots corresponding to the green and red circles in Fig. 5. Climatology plots for SWE and water year-to-date precipitation (PCP) are shown in units of millimeters. Climatology plots for SM2 and SM3 are shown in units of volumetric water content.
Citation: Journal of Hydrometeorology 14, 5; 10.1175/JHM-D-12-0136.1
Note that the stations used as predictors in the NRCS forecast models are a subset of all the existing stations in the vicinity of the basin. This subset was selected by NRCS hydrologists to maximize forecast skill, discarding predictor data from other stations because they did not improve forecast performance. The results for BMDC2 also illustrate two limitations of our approach, which are that the locations of existing stations and issues of scale between the grid and the point element are not explicitly accounted for in the selection methodology. It is possible that the variances of some gridded/simulated data are different enough from those of collocated SNOTEL data that the former improves forecast skill while the latter does not (or vice versa). Forecast improvements in grid cells occupied by existing stations sometimes occurred, particularly in basins with dense existing networks.
A different scenario is presented for February forecasts at HRDM8 in Fig. 5. This is an example of a relatively ungauged basin, where NRCS relies on data from nearby stations for predictors in their current forecast models. Improvements tend to increase toward the northwesterly direction, with the largest occurring in the same grid cell as the basin outlet. The corresponding plots in Fig. 6 indicate that these improvements are due almost entirely to precipitation, with snow-covered area accounting for just a small percentage of the basin at its southern tip. As in the prior example, the grid cell identified as best is located in an area of both high correlation and variability and, additionally, in a region of low intercorrelation with the other predictors in the forecast model. For March forecasts at MONO3, the selected locations also occupy an ungauged area of the basin. Figure 6 indicates that these improvements are due to both SWE and precipitation, suggesting that the ability to exploit the predictive power of multiple variables is yet another reason for selection.
Figure 5 also presents examples of forecast improvements from EXP2, which, as depicted by the sizes of these circles, are larger than those for EXP1. Other commonalities include the forecast months, which are at various points in the accumulation season, and the locations of the grid cells offering improvements, which occur primarily in the basin valleys. The predictor patterns underlying these improvements (Fig. 6), however, are slightly different for each basin. For BITM8, the dominant two predictor types are SWE and SM3 (the deeper, larger soil layer). Both best-overall and month-optimized locations are found in areas of high variability and somewhat higher correlations for each predictor type. One striking result is the low correlation for SM3 at the higher elevations along the basin's periphery, despite these locations being generally wetter than those in the center. This pattern occurs again at DNRC2, where the largest improvements seem to follow the river's course. For GBRW4, SM2 and SM3 combine to offer the largest forecast improvements, also in the lowlands of the basin. Inspection of soil types for areas offering improvements from soil moisture predictors revealed no consistent patterns. For BITM8, DNRC2, and GBRW4, for example, predominant soil types include sandy loam, clay loam, and silt loam, respectively, at the depth of the third soil layer (Miller and White 1998). Note that we found SNOTEL installations to be generally unsupported, for purposes of seasonal streamflow forecasting, in wilderness areas; for HHWM8, for example, we found little to no improvement in forecast skill, and in other basins, comparable improvements were found both within and outside wilderness boundaries.
c. Snow cover validation and predictor contributions
Figure 7 presents results from the snow cover comparisons between satellite and simulated data. Since, as noted above, many of the locations selected in the network design analysis were associated with SWE predictors at the fringes of annual snowpack, comparisons for these sites acquired additional importance. Agreement varied from basin to basin, but generally indicated a binary match percentage of about 80%.
Results of binary snow cover comparisons between VIC and MODIS for those grid cells that were selected as best overall in the network design analysis. The numbers above each bar represent the number of forecast months for which comparisons were performed.
Citation: Journal of Hydrometeorology 14, 5; 10.1175/JHM-D-12-0136.1
An examination of predictor contributions for various forecast models and basins reveals some interesting patterns (Fig. 8). When only the standard predictors are considered (EXP1), the contribution of precipitation is generally large, while the influence of SWE tends to increase as the year progresses, depending on the basin. When soil moisture is included (EXP2), however, the contribution of precipitation is diminished, while the influence of soil moisture appears stronger in the early part of the water year and weakens as the year progresses. These results are similar to the improvements in forecast skill for EXP1 and EXP2, which tended to be greater during the ablation season and accumulation season, respectively, and consistent with expectations for the western U.S. hydrologic cycle (Wood and Lettenmaier 2008). Interestingly, the contribution of soil moisture seems more significant in some of the semiarid basins than wetter basins where precipitation may play a larger predictive role. Furthermore, soil moisture predictors are mostly from the bottom two soil layers during the accumulation season, transitioning to the top two layers during the ablation season. Note that all precipitation predictors are water year-to-date, and that, for some of the basins, streamflow observations are used in the NRCS forecast models as described in section 3a(1). Also note the relatively strong presence of observed predictors even when the vast majority of the predictor candidates are gridded or simulated, as they are in the columns labeled “All Grid Cells.”
Percentage of seasonal streamflow variance that is explained by each predictor for forecast models labeled at top. Hatched subsections denote gridded/simulated predictors, while those that are solid denote observed predictors in the NRCS forecast models. For months without gridded/simulated predictors, no improvement over baseline skill was found. Note that, for forecast models in the “Best Overall” columns, it is possible to have multiple gridded/simulated predictors if more than one predictor type was selected for the best overall grid cell.
Citation: Journal of Hydrometeorology 14, 5; 10.1175/JHM-D-12-0136.1
d. Operational analysis
Skill results for the reformulated forecast models are shown for six of the seven Upper Colorado study basins in Fig. 9. The shaded areas represent the range of skill that can be expected by adding a predictor to the NRCS forecast models, with the lower bounds determined as described in section 3e and the upper bounds equivalent to the lines representing the best grid cell overall in Fig. 4. Given the subjective nature of the NRCS forecast model selection process, we expect actual skill to fall somewhere in between. Comparisons of these plots with those representing NWS forecast skill reveal a mixture of results. In two of the basins (BMDC2 and NVRN5), current NRCS skill is either comparable to or better than that of NWS, and an additional predictor is expected to further improve NRCS skill. In the other four basins, current NWS/ESP forecasts show slightly greater skill than their NRCS counterparts, but predictor data from a new SNOTEL station seem to equalize their performance, perhaps because of the inclusion of new water balance variables as predictors.
Skill of various operational forecasts for six of the seven basins in the Upper Colorado water resource region. Purple NRCS lines correspond to those in Fig. 3. Shaded blue and brown areas indicate expected skill for NRCS forecasts that incorporate data from an additional observation station under the EXP1 and EXP2 scenarios, respectively.
Citation: Journal of Hydrometeorology 14, 5; 10.1175/JHM-D-12-0136.1
5. Discussion and conclusions
We have demonstrated a skill-oriented, hybrid dynamical–statistical methodology to inform the expansion of hydrometric networks for statistical seasonal streamflow forecasts. While the approach was developed and tested in the western United States, it is appropriate for any setting in which seasonal streamflow forecast skill is strongly influenced by IHCs. Similarly, the foundation of the approach can be generalized to other water resources applications involving the use of point observations for statistical prediction models, such as those involving groundwater networks or water quality networks.
Evaluation of the method in the western United States revealed that locations identified as optimal for SNOTEL placement are primarily concentrated in regions of high correlation with seasonal streamflow, with additional commonalities including high predictor variability, low cross correlation with existing predictor data, and the ability to exploit the predictive power of multiple water balance variables. When only SWE and precipitation predictors are considered, these tend to occur at the margins of the average snowpack for a given forecast month, though resulting improvements in skill are only notable for basins with sparse existing networks or late in the snow ablation season. One can speculate that the mechanism behind this relationship is related to a second mode of interannual variability that is not captured by existing predictors with higher climatological averages. Another possibility involves the concept of the snow depletion curve, with the binary snow cover signal at the identified location implicitly providing some indication of snow covered area in the basin and, consequently, the mean areal SWE of the snowpack.
When soil moisture is added as a predictor, improvements in skill are more significant. As above, the mechanism responsible for this result is potentially related to an uncorrelated mode of interannual or lower-frequency variability; the power spectrum of soil moisture, for example, has been shown to exhibit a strong coherence to that of indicators of climatic extremes such as drought (Lakshmi et al. 2004). The largest improvements are found during the accumulation season, when selected locations are typically concentrated in low to midelevations, transitioning to higher elevations as the water year progresses. These patterns are likely also related to snow cover, which shields soil from moisture variations at higher elevations where winter precipitation falls entirely as snow and no winter melt occurs. In contrast, soil moisture at lower elevations, where soil profiles are typically deeper and water storage capacity greater, is more active and better able to reflect the degree of water year-to-date precipitation. As the melting snow exposes the underlying ground surface, the once-dormant high-elevation soil moisture is again altered in relation to the snowpack volume.
Several recent studies have focused on characterizing the importance of IHCs (Maurer and Lettenmaier 2003; Wood and Lettenmaier 2008), and soil moisture in particular (Koster et al. 2010; Mahanama et al. 2012), for water supply forecasts. Mahanama et al. (2012) found that, outside of spring, the impact of soil moisture initialization on ensemble forecasts dominated over that of snow initialization, and fall soil moisture initialization contributed to skill at particularly long leads. This study shows that such soil moisture predictability can be harnessed in operational statistical water supply forecasts, particularly during the accumulation season using (in most basins) sensors at low to midelevations. For many basins, even a single soil moisture predictor can enhance forecast skill, and the approach described herein can be used to rank basins in order of those with the most to gain.
The distinction between observed and modeled data represents an important issue in this study. One potential reason for the nominal improvements found for EXP1 is that the gridded precipitation data used in the analysis are partially based on the same SNOTEL observations already in the NRCS statistical models [see section 3b(2)]. At the same time, it is uncertain whether the greater forecast improvements found for EXP2 can be expected from in situ observations. Correlations between simulated and observed soil moisture are generally inconsistent, possibly because of fundamental differences between these two quantities; Koster et al. (2009), for example, have described simulated soil moisture as a “model-specific index of wetness” with no direct observational analog. Nonetheless, the strong presence of observed predictors in forecast models for which the majority of predictor candidates are simulated or gridded (section 4c) suggests that soil moisture observations should be at least as useful as proxy simulated data for statistical forecasts, although at present, their short record lengths preclude testing this hypothesis.
An additional finding of this research is that simulated hydrologic data can also be combined with observations to improve operational statistical water supply forecasts, a strategy that may prove more effective, and is almost certainly less expensive, than network augmentation in the near term. Indeed, an interesting question is whether advances in computing power and numerical models will render investments in new observations less worthwhile from a forecasting perspective than those in simulation model-based (and ensemble-based) prediction methods. This research demonstrates that, at present, statistical forecasts are comparable in skill to model-based forecasts, and synergies result from their combination; future work involving ensemble-based methods may benefit from similar comparisons with dynamical–statistical approaches. Nonetheless, forecast benefits resulting from SNOTEL installations today are difficult to realize in a statistical framework until enough time has elapsed to develop a statistical climatology. The effects of climate change and nonstationarity on hydrologic forecast methods are also a relevant topic (Milly et al. 2008; Wood 2007; Brekke et al. 2010). We suggest, however, that well-placed observations provide important indications of actual conditions in any climate and that statistical forecasts will remain useful both for their ability to capture linear predictability at relatively low cost and as benchmarks against which to evaluate the skill of more intensive prediction approaches.
Acknowledgments
This research was supported by National Oceanic and Atmospheric Administration Grant NA11OAR4310150. We sincerely thank the staff of NRCS NWCC for their generous assistance, with special thanks to David Garen, Tom Perkins, Gus Goodbody, Rashawn Tama, Jolyne Lea, and Cara McCarthy. The thoughtful comments of Jessica Lundquist and three anonymous reviewers are also gratefully acknowledged.
REFERENCES
Boardman, H. P., 1936: The effect of soil-absorption on snow-survey forecasting of stream-flow. Trans., Amer. Geophys. Union,17, 534–537.
Brekke, L. D., Garen D. , Werner K. , and Laurine D. , 2010: Projecting climate change impacts on seasonal water supply forecasting error. Preprints, 18th Conf. on Applied Climatology/22nd Conf. on Climate Variability and Change/24th Conf. on Hydrology, Atlanta, GA, Amer. Meteor. Soc., J15.4. [Available online at https://ams.confex.com/ams/90annual/techprogram/paper_162386.htm.]
Clyde, G. D., 1940: Soil-moisture studies as an aid in forecasting runoff from snow-cover. Trans., Amer. Geophys. Union,21, 871–873.
Day, G. N., 1985: Extended streamflow forecasting using NWSRFS. J. Water Resour. Plann. Manage., 111, 157–170, doi:10.1061/(ASCE)0733-9496(1985)111:2(157).
Fiering, M. B., 1965: An optimization scheme for gaging. Water Resour. Res., 1, 463–470, doi:10.1029/WR001i004p00463.
Garen, D. C., 1992: Improved techniques in regression-based streamflow volume forecasting. J. Water Resour. Plann. Manage., 118, 654–670, doi:10.1061/(ASCE)0733-9496(1992)118:6(654).
Garen, D. C., and Pagano T. C. , 2007: Statistical techniques used in the VIPER water supply forecasting software. NRCS-USDA Engineering-Snow Survey and Water Supply Forecasting Tech. Note 210-2, 18 pp. [Available online at http://www.wcc.nrcs.usda.gov/ftpref/downloads/factpub/wsf/technotes/Tech_note_statistical_techniques_in_Viper.pdf.]
Hall, D. K., Riggs G. A. , and Salomonson V. V. , 2006: MODIS/Terra Snow Cover 8-Day L3 Global 0.05deg CMG Version 5, Oct 2000 to Nov 2010. National Snow and Ice Data Center, Boulder, CO, digital media. [Available online at http://nsidc.org/data/mod10c2.html.]
Hartman, R. K., and Henkel A. F. , 1994: Modernization of statistical procedures for water supply forecasting. Proc. 62nd Annual Western Snow Conf., Santa Fe, NM, Western Snow Conference, 104–114. [Available online at http://www.westernsnowconference.org/sites/westernsnowconference.org/PDFs/1994Hartman.pdf.]
Helms, D., Phillips S. E. , and Reich P. F. , Eds., 2008: The History of Snow Survey and Water Supply Forecasting: Interviews with U.S. Department of Agriculture Pioneers. U.S. Department of Agriculture, 306 pp.
Koster, R. D., Guo Z. , Yang R. , Dirmeyer P. A. , Mitchell K. , and Puma M. J. , 2009: On the nature of soil moisture in land surface models. J. Climate, 22, 4322–4335, doi:10.1175/2009JCLI2832.1.
Koster, R. D., Mahanama S. P. P. , Livneh B. , Lettenmaier D. P. , and Reichle R. H. , 2010: Skill in streamflow forecasts derived from large-scale estimates of soil moisture and snow. Nat. Geosci., 3, 613–616, doi:10.1038/ngeo944.
Lakshmi, V., Piechota T. , Narayan U. , and Tang C. , 2004: Soil moisture as an indicator of weather extremes. Geophys. Res. Lett., 31, L11401, doi:10.1029/2004GL019930.
Landres, P., Alderson J. , and Parsons D. J. , 2003: The challenge of doing science in wilderness: Historical, legal, and policy context. George Wright Forum, Vol. 20 (3), George Wright Society, 42–49. [Available online at http://www.fs.fed.us/rm/pubs_other/rmrs_2003_landres_p001.pdf.]
Lea, J., and Harms D. , 2011: Developing NRCS SNOTEL and SCAN soil moisture parameters for water supply forecasting applications. Proc. 79th Annual Western Snow Conf., Stateline, NV, 109–112. [Available online at http://www.westernsnowconference.org/sites/westernsnowconference.org/PDFs/2011Lea.pdf.]
Liang, X., Lettenmaier D. P. , Wood E. F. , and Burges S. J. , 1994: A simple hydrologically-based model of land surface water and energy fluxes for general circulation models. J. Geophys. Res., 99, 14 415–14 428, doi:10.1029/94JD00483.
Livneh, B., Rosenberg E. A. , Lin C. , Nijssen B. , Mishra V. , Andreadis K. M. , Maurer E. P. , and Lettenmaier D. P. , 2013: A long-term hydrologically based dataset of land surface fluxes and states for the conterminous United States: Update and extensions. J. Climate, doi:10.1175/JCLI-D-12-00508.1, in press.
Mahanama, S., Livneh B. , Koster R. , Lettenmaier D. , and Reichle R. , 2012: Soil moisture, snow, and seasonal streamflow forecasts in the United States. J. Hydrometeor., 13, 189–203, doi:10.1175/JHM-D-11-046.1.
Maurer, E. P., and Lettenmaier D. P. , 2003: Predictability of seasonal runoff in the Mississippi River basin. J. Geophys. Res., 108, 8607, doi:10.1029/2002JD002555.
Maurer, E. P., Wood A. W. , Adam J. C. , and Lettenmaier D. P. , 2002: A long-term hydrologically based dataset of land surface fluxes and states for the conterminous United States. J. Climate, 15, 3237–3251, doi:10.1175/1520-0442(2002)015<3237:ALTHBD>2.0.CO;2.
McCuen, R. H., 1985: Statistical Methods for Engineers. Prentice-Hall, 439 pp.
Miller, D. A., and White R. A. , 1998: A conterminous United States multilayer soil characteristics data set for regional climate and hydrology modeling. Earth Interact., 2, doi:10.1175/1087-3562(1998)002<0001:ACUSMS>2.3.CO;2.
Milly, P. C. D., Betancourt J. , Falkenmark M. , Hirsch R. M. , Kundzewicz Z. W. , Lettenmaier D. P. , and Stouffer R. J. , 2008: Stationarity is dead: Whither water management? Sci., 319, 573–574, doi:10.1126/science.1151915.
Mishra, A. K., and Coulibaly P. , 2009: Developments in hydrometric network design: A review. Rev. Geophys., 47, RG2001. doi:10.1029/2007RG000243.
Molotch, N. P., and Bales R. C. , 2005: Scaling snow observations from the point to the grid element: Implications for observation network design. Water Resour. Res., 41, W11421, doi:10.1029/2005WR004229.
Morin, G., Fortin J. P. , Sochanska W. , Lardeau J. P. , and Charbonneau R. , 1979: Use of principal component analysis to identify homogenous precipitation stations for optimal interpolation. Water Resour. Res., 15, 1841–1850, doi:10.1029/WR015i006p01841.
Pagano, T. C., Garen D. C. , Perkins T. R. , and Pasteris P. A. , 2009: Daily updating of operational statistical seasonal water supply forecasts for the western U.S. J. Amer. Water Resour. Assoc., 45, 767–778, doi:10.1111/j.1752-1688.2009.00321.x.
Pardo-Igúzquiza, E., 1998: Optimal selection of number and location of rainfall gauges for areal rainfall estimation using geostatistics and simulated annealing. J. Hydrol., 210, 206–220, doi:10.1016/S0022-1694(98)00188-7.
Peck, E. L., and Schaake J. C. , 1990: Network design for water supply forecasting in the West. J. Amer. Water Resour. Assoc., 26, 87–99, doi:10.1111/j.1752-1688.1990.tb01354.x.
Perkins, T. R., Marron J. K. , and Goodbody A. G. , 2010: ArcGIS technique to evaluate the SNOTEL data network. Proc. Second Joint Federal Interagency Conf., Las Vegas, NV, 11 pp. [Available online at http://acwi.gov/sos/pubs/2ndJFIC/Contents/1F_tperkins.pdf.]
Rodríguez-Iturbe, I., and Mejía J. M. , 1974: The design of rainfall networks in time and space. Water Resour. Res., 10, 713–728, doi:10.1029/WR010i004p00713.
Rosenberg, E. A., Wood A. W. , and Steinemann A. C. , 2011: Statistical applications of physically based hydrologic models to seasonal streamflow forecasts. Water Resour. Res., 47, W00H14, doi:10.1029/2010WR010101.
Schaefer, G. L., Cosh M. H. , and Jackson T. J. , 2007: The USDA Natural Resources Conservation Service Soil Climate Analysis Network (SCAN). J. Atmos. Oceanic Technol., 24, 2073–2077, doi:10.1175/2007JTECHA930.1.
Speers, D. D., Rockwood D. M. , and Ashton G. D. , 1996: Snow and snowmelt. Hydrology Handbook, 2nd ed. ASCE Manuals and Reports on Engineering Practice, No. 28, American Society of Civil Engineers, 437–476.
Tangborn, W. V., 1980: A model to forecast short-term snowmelt runoff using synoptic observations of streamflow, temperature, and precipitation. Water Resour. Res., 16, 778–786, doi:10.1029/WR016i004p00778.
Tsintikidis, D., Georgakakos K. P. , Sperfslage J. A. , Smith D. E. , and Carpenter T. M. , 2002: Precipitation uncertainty and raingauge network design within Folsom Lake watershed. J. Hydrol. Eng., 7, 175–184, doi:10.1061/(ASCE)1084-0699(2002)7:2(175).
Volkmann, T. H. M., Lyon S. W. , Gupta H. V. , and Troch P. A. , 2010: Multicriteria design of rain gauge networks for flash flood prediction in semiarid catchments with complex terrain. Water Resour. Res., 46, W11554, doi:10.1029/2010WR009145.
Wood, A. W., 2007: The effects of climate change on water supply forecasting in the Feather River basin. Preprints, Fourth Annual California Climate Change Conf., Sacramento, CA, California Energy Commission.
Wood, A. W., and Lettenmaier D. P. , 2008: An ensemble-based approach for the attribution of streamflow prediction uncertainty. Geophys. Res. Lett., 35, L14401, doi:10.1029/2008GL034648.
