Wind-induced losses, or undercatch, can have a substantial impact on precipitation gauge observations, especially in alpine environments that receive a substantial amount of frozen precipitation and may be exposed to high winds. A network of NOAH II all-weather gauges installed in the Snowy Mountains since 2006 provides an opportunity to evaluate the magnitude of undercatch in an Australian alpine environment. Data from two intercomparison sites were used with NOAH II gauges with different configurations of wind fences installed: unfenced, WMO standard double fence intercomparison reference (full DFIR) fences, and an experimental half-sized double fence (half DFIR). It was found that average ambient temperature over 6-h periods was sufficient to classify the precipitation phase as snow, mixed precipitation, or rain in a statistically robust way. Empirical catch ratio relationships (i.e., the quotient of observations from two gauges), based on wind speed, ambient temperature, and measured precipitation amount, were established for snow and mixed precipitation. An adjustment scheme to correct the unfenced NOAH II gauge data using the catch ratio relationships was cross validated with independent data from two additional sites, as well as from the intercomparison sites themselves. The adjustment scheme was applied to the observed precipitation amounts at the other sites with unfenced NOAH II gauges. In the worst-case scenario, it was found that the observed precipitation amount would need to be increased by 52% to match what would have been recorded had adequate shielding been installed. However, gauges that were naturally well protected, and those below about 1400 m, required very little adjustment. Spatial analysis showed that the average seasonal undercatch was between 6% and 15% for gauges above 1000 m MSL.
Precipitation in the Australian alpine regions is critically important to agriculture in the Murray–Darling basin, which accounts for 45% of national irrigated production (Worboys and Good 2011). Alpine rain and snowfall also support major hydroelectricity infrastructure and a ski tourism industry. A number of vulnerable species are threatened by decreases in annual snowpack levels (Nicholls 2005) and projected decreases in precipitation (Hennessy et al. 2003). Despite these factors, precipitation observations in both the Snowy Mountains alpine region (New South Wales; see Fig. 1) and in the Victorian Alps are sparse and of poor reliability, with long-term precipitation measurements limited to just a few locations near ski resorts.
The climate of the Snowy Mountains received some attention following the Millennium Drought (van Dijk et al. 2013) in southeastern Australia. It was the subject of a study by Chubb et al. (2011), who performed a decomposition of wintertime precipitation in 1990–2009 by synoptic type. Precipitation was predominantly due to cold fronts associated with either cutoff lows (equatorward of 45°S) or lows embedded in the westerly storm track, in roughly equal proportions. They noted that the topography was of principal importance in determining the spatial impacts of the drought, with declines on the western (windward) slopes of the mountains related much more closely to elevation than on the eastern slopes. By way of explanation, Fiddes et al. (2015) noted that the synoptic types affecting eastern slopes precipitation were less influenced by the climate drivers of the Millennium Drought. Dai et al. (2013) also identified synoptic types associated with precipitation, but they were based on a nearby upwind sounding. Two of the classes accounted for more than 70% of the total wintertime precipitation and were typified by high moisture flux and/or high shear.
Accurate measurement of precipitation in mountainous regions is extremely challenging. High spatial variability due to orographic effects necessitates the use of high-density precipitation gauge networks to accurately represent precipitation patterns (Frei and Schär 1998). A further complication is that different gauge configurations can record quite different amounts at the same location during snow or mixed precipitation (Goodison et al. 1981, 1997; Rasmussen et al. 2012), leading to artifacts in records from inhomogeneous networks.
Undercatch (Groisman and Legates 1994), or losses due to wind, have long been recognized as an error source for frozen precipitation measurement due to flow deformation around the gauge orifice. Wild (1885) observed that a gauge shielded by a fence collected substantially more snow than an unshielded one, and Alter (1937) was among the first to test the collection efficiency of a variety of shield designs. His legacy persists in the contemporary use of the Alter shield, which consists of a ring of wedge-shaped slats surrounding the gauge and suspended slightly above the gauge orifice. It has since been shown that a gauge shielded in this way can catch up to 50% more precipitation than an unshielded counterpart (Goodison et al. 1981), but this value depends heavily on the type of gauges used and the conditions in which they are operating. Double fences were found to be even more effective; Golubev (1986) found that the WMO standard double fence intercomparison reference (DFIR) gauge, consisting of a Russian Tretyakov gauge housed in concentric 12- and 6-m diameter octagonal fences [see Goodison et al. (1997), annex 2B], recorded 92%–96% of the reference bush gauge amount. Yang (2014) recently compared 20 years of DFIR and bush gauge data collected over 12-h intervals to calculate correction equations to recover the true precipitation amount from DFIR gauge observations.
There is continued interest in gauge intercomparison research. A number of national standard precipitation gauges were assessed in the WMO Solid Precipitation Measurement Intercomparison [Goodison et al. (1997), and references therein]. One of the main aims of this project was to establish standard methods of adjusting solid precipitation measurements for gauge undercatch. The dominant environmental variable affecting gauge efficiency was found to be wind speed (measured at gauge orifice height), and temperature was found to be more important during mixed precipitation because of higher snowflake fall speeds as melting occurs. More recently, Rasmussen et al. (2012) highlighted the ongoing disparity in the performance of different gauges, and a second WMO experiment [the WMO Solid Precipitation Intercomparison Experiment (SPICE; Nitu et al. 2012)] is currently being conducted at a number of sites around the world, including Guthega Dam (GD) in the Australian Snowy Mountains, which is the focus of this paper.
A number of factors affect precipitation particle fall speed and thus the gauge collection efficiency. Rain and mixed precipitation (wet snow) typically have higher fall speeds than dry snow (Yuter et al. 2006). Thériault et al. (2012) directly simulated the flow around a Geonor gauge in a single Alter shield and computed collection efficiencies for wet (fall speed ≥−2 m s−1) and dry (fall speed ≃ 1m s−1) snowflakes and compared these predictions with observations of snow crystal types during a wintertime storm in Colorado. The wet crystal category, in fact consisting of a wide range of crystal types, was found to have a higher scatter in the collection efficiency because of the wider range of particle fall speeds.
Other losses that influence the amount of precipitation recorded are wetting losses, which are systematic undermeasurements when volumetric methods (i.e., tipping gathered precipitation into a separate measuring container) are used, and evaporative losses, which occur in the interval between bucket tips or manual gauge readings. Trace precipitation occurs when the accumulated amount is less than the increment of the gauge. For manual observations, this does not contribute to the measured seasonal precipitation amount but may nonetheless be important in low-precipitation regimes.
Determination of the phase of the precipitation generally requires either manual observation of the gauge contents or specialized equipment, so in practice it is often predicted from meteorological variables. The algorithms used may be basic temperature thresholds for snow, mixed precipitation, and rain, as used by Yang and Ohata (2001). The wet-bulb temperature is a better indication of conditions in the immediate vicinity of a precipitation particle, which cools both itself and the air around it through evaporation and/or sublimation, and Michelson (2004) evaluated the mixed fraction for −0.21° < < 2.42°C using hyperbolic functions. Harder and Pomeroy (2013) estimated the hydrometeor temperature using psychometric energy balance (which also requires a humidity measurement) to determine precipitation phase. They found appreciable increases in the accuracy of phase discrimination compared to temperature-only schemes, especially as the time interval used decreased from daily to subhourly. More complex algorithms use aerological data to identify the presence of warm layers where snowflake melting could occur (e.g., Bourgouin 2000) in order to predict the occurrence of freezing rain, but this approach requires intensive soundings and/or numerical model data.
Once the relationship [often termed the catch ratio (CR)] between a reference precipitation gauge and the gauges used in the field have been established, it is possible to assess the impact of environmental losses over a region and develop an estimate for the seasonal precipitation amounts. A number of studies have been performed around the globe with this objective, with highly variable results due to the differences in the nature of the sites and equipment used:
In Siberia, application of the regression parameters estimated by the WMO intercomparison showed that annual precipitation was underestimated by 10%–65% (Yang and Ohata 2001).
Mongolian annual precipitation amounts should be 17%–42% greater than reported from 31 meteorological stations because of undercatch of snowfall (Zhang et al. 2004).
In Alaska, 10 gauges were adjusted by 10%–140% over a 2-yr period to account for wind-induced losses (Yang et al. 1998).
Chinese gauge data were adjusted at 710 meteorological stations during 1951–98 by Ye et al. (2004), identifying an underestimation of 6%–62% (mean 19%) for annual precipitation over all gauges.
Adjustment of a long-term precipitation dataset for gauges north of 45°N from multiple countries required correction factors of 80%–120% in winter months and around 10% in summer months (Yang et al. 2005).
In this paper, we seek to test two key hypotheses. The first is that a precipitation phase partitioning scheme can be developed based on the observations available at a gauge intercomparison site in the Snowy Mountains region. This site did not have humidity measurements for the majority of the analysis period, so we used a temperature-only scheme and tested it using the catch ratio between an unfenced (UF) and a fenced gauge. The second major hypothesis is that an empirical CR relationship between fenced and unfenced gauges can be modeled by observations from the site and that this relationship can be used to adjust precipitation amounts reported by an unfenced gauge in a way that reduces the root-mean-square error (RMSE) between the two gauges. We aim to validate this adjustment scheme using independent data before applying it across the network to evaluate the impact of installing wind fences throughout the Snowy Mountains region.
Similar analysis has been presented in other studies, but this paper is the first to present such work for an Australian alpine environment, and there are several aspects of this work that are of general interest. First, to our knowledge, the use of catch ratios to evaluate the temperature-only partitioning is novel and the analysis is compelling. Further, we show that the inclusion of a precipitation term in the CR formulation, which is also a novel approach, results in an improved fit and substantially reduces the RMSE in the adjustment scheme.
2. Surface meteorological observations in the Snowy Mountains region
Snowy Hydro Ltd. (SHL) is the operator of the Snowy Mountains Hydro-electric Scheme and has maintained an independent surface meteorological network to inform hydrological inflow estimates. To facilitate the evaluation of the Snowy Precipitation Enhancement Research Project (SPERP; Manton et al. 2011), there have been substantial improvements to the observational network.
The data were quality coded and provided as half-hourly accumulated values for precipitation, and half-hourly average values for other variables, as soon as practicable after collection. In this paper, we use averages and totals over 6-h intervals that divide each day into four parts, for an analysis period of the months of April–October in the years of 2007–12.
a. Precipitation gauges
Figure 1 (right) shows the current precipitation gauge network, which has NOAH II (ETI Instrument Systems Inc. 2008) at locations where snow frequently falls, and a mixture of heated and unheated tipping-bucket gauges (TBG) throughout the region. All of the NOAH II gauges in the Snowy Mountains were fitted with the manufacturer’s Alter shield. At many of these sites other variables are observed, including temperature, wind speed and direction, atmospheric pressure, and humidity.
1) NOAH II precipitation gauges
The NOAH II gauge collects precipitation in a chamber partially filled with antifreeze, while the pressure at the bottom is monitored by a sensitive load cell. These data were processed by a software algorithm to generate electrical pulses in real time similar to a tipping-bucket rain gauge. The raw data were subsequently reprocessed using a similar algorithm by SHL staff as follows to provide the half-hourly dataset used in this paper. The raw fluid depth and transducer temperature data were median filtered with a centered 50-min window to eliminate noise and other artifacts. The filtered fluid depth is periodically compared to a reference depth, and if it is greater by one (or more) increments of 0.254 mm, a corresponding number of tips are reported, and the reference depth is incremented. If the fluid depth falls below the reference depth for a sustained period (e.g., through evaporation), the reference depth is correspondingly reduced.
In addition to the Alter shields, a large number of the NOAH II gauges were protected from wind by a double fence. These fences were made to half of the diameter of the standard WMO DFIR fence (hereafter full and half DFIR fences), with the outer fence inscribed in a 6-m diameter circle, to facilitate prefabrication and transportation by helicopter, as well as to reduce the environmental impact within the national park.
Wetting losses are reduced by the application of a hydrophobic coating to the wall of the gauge. However, there is a significant surface area (especially when the fluid level is low) and some retention and evaporation of moisture is possible. To suppress evaporation, a layer of oil was maintained above the antifreeze. This technique, combined with the software algorithms, prevented virtually all evaporative losses. Trace precipitation, on the other hand, could fail to be carried over to the next precipitation event because of the temperature dependence of the load cell. Changes in temperature due to the diurnal cycle were sufficient to cause the software algorithm to reset the reference depth periodically, so that partial increments could be lost. This response was expected to be similar to that experienced because of evaporation by tipping-bucket gauges.
2) Tipping-bucket gauges
There were a number of different configurations of tipping-bucket gauges, with 37 installed in total throughout the Snowy Mountains region. Some of these are installed inside heated chambers, others have heated collectors, and others at lower elevations were not heated at all. In addition to undercatch, these gauges were subject to wetting and evaporative losses, as well as reporting delays due to buildup of snow in the collectors, but treatment of these impacts is beyond the scope of this paper. For the spatial analysis performed in section 2c, we only used tipping-bucket gauges below 1400 m to minimize this effect.
b. Other surface meteorological variables
In addition to the precipitation gauges, a number of other surface meteorological instruments were deployed. At most of the sites, at least wind speed and direction and ambient temperature were recorded in addition to precipitation amount. A smaller subset of the sites also reported pressure and humidity, but these variables were not used in this study.
For the SPERP, at sites where power was available, heated NRG IceFree3 cup anemometers and wind vanes were deployed, and at other sites either Young 05103 or Vaisala WMS301 combination wind monitors were used. A number of other instruments from the original SHL sites were incorporated into the SPERP dataset and have been used in this study. In total, wind observations were reported at 43 unique sites during the analysis period.
Ambient temperature was reported by Unidata 6507A thermistor sensors at the majority of the new SPERP sites, but Vaisala HMP45A humidity–temperature sensors were deployed at a small number of sites instead. Some of the original SHL sites used in-house thermocouple sensors, and others had Vaisala HMP131Y humidity–temperature sensors. In total, 47 unique sites reported temperature during the analysis period.
c. Guthega Dam intercomparison site
The observational site near Guthega Dam (elevation 1586 m MSL) is the principal subject of this paper and has since been included in the ongoing WMO SPICE (Nitu et al. 2012). It is situated on a small plateau about 5 m above the level of the dam (see Fig. 2). The small bluff indicated in Fig. 2 is about 2 m in height and is not expected to cause substantial vertical wind at the gauge orifices, but these cannot be ruled out with the available observations. Since 2006, there have been four precipitation gauges at Guthega Dam (see Table 1), including three NOAH II gauges (full DFIR fence, half DFIR fence, and unfenced), and a Hydrological Services (HS) tipping-bucket gauge installed in a heated chamber. Precipitation (Pr > 0.254 mm) was recorded in 35% of all 6-h intervals during the analysis periods. The mean (nonzero) precipitation amount was 4.76 mm, and the median was 2.03 mm.
Two NRG IceFree3 anemometers were installed outside the wind fences at Guthega Dam at 3 m above ground level, at the same height as the top of the NOAH II gauges, and two others were installed within the fence structures. These fenced anemometers consistently reported lower wind speeds by a factor of 4–5 than their unfenced counterparts, demonstrating the efficacy of the DFIR fences. We designated the wind speed variables as per the labels in Fig. 2: the wind speed from the anemometer installed beside the half DFIR wind fence is designated WSA, and the one beside the unfenced NOAH II gauge is WSB. Because of maintenance and certain instrumentation issues, each of these was available only about 60% of the time over the six winters. To establish a more complete and reliable wind speed record for the observation period, a composite of these wind speeds, designated WS, was used. WS was available for 82% of the 6-h observation intervals in the analysis periods of April–October 2007–12 and was validated against the independent wind speed WSD measured at 10 m.
On average, winds came from the western sector at Guthega Dam with an average speed of about 3.2 m s−1 (see Figs. 3a,b). When precipitation was recorded, the winds predominantly came from the northwest, with an average speed of about 4.3 m s−1. Wind speed was weakly correlated with precipitation amount, with maximum precipitation rates observed for wind speeds of around 5 m s−1.
A composite of data from the two thermometers at the site provided a robust ambient temperature T that was available for more than 99% of the 6-h averaging intervals. The mean temperature over all the 6-h measurement intervals was 2.8°C, and for those in which precipitation was recorded the mean temperature was 2.3°C.
d. The Kerries observational site
At the Kerries (TK), a site about 13 km to the north of Guthega Dam at 1741 m MSL elevation, there were two NOAH II gauges installed simultaneously for four of the winters of the analysis period (see Fig. 2, Table 1). One of these gauges was unfenced and was removed at the end of October 2010, and the other was deployed for the entire period inside a half DFIR fence. There was also a heated tipping-bucket configuration similar to the one at Guthega Dam, but this was unused in our study. In terms of precipitation amount statistics, the site was almost identical to Guthega Dam, with precipitation recorded in 34% of nominal observation periods, with a mean (nonzero intervals) of 4.34 mm and median of 2.03 mm.
The Kerries was considerably better protected than Guthega Dam. Mean wind speeds at the site were 1.5 m s−1 for all intervals and 1.7 m s−1 for intervals with precipitation. The easterly sectors in Figs. 3c and 3d were probably due to nocturnal drainage flow down a gully running in this direction just to the north of the site and were generally not present during intervals with precipitation. The mean temperature over all the 6-h measurement intervals was 1.0°C, and for those in which precipitation was recorded the mean temperature was 0.2°C.
e. Other sites with multiple gauges
Two other sites had a period with unfenced and half DFIR fenced configurations of NOAH II gauges installed simultaneously, which we used to cross validate the CR models developed in this paper. Grey Hill, to the west of the Main Range in Fig. 1 at an elevation 1624 m MSL, had 162 intervals where all data were available and precipitation was recorded. Snowy Plain, to the east at 1368 m MSL, had 226 nominal intervals. Both of these sites were relatively protected with mean wind speeds around 2 m s−1 during intervals where precipitation was recorded.
f. Gap filling and spatial interpolation
In a number of cases, wind speed and/or temperature instrumentation were not installed until well after the installation of the NOAH II gauges. Additionally, there are a number of extended gaps in the datasets where instrument maintenance was delayed because of logistical difficulties (many of these sites were only accessible by helicopter in the winter months).
For the catch ratio analysis and evaluation in sections 4 and 5, only directly measured variables were used. To apply the results of our analysis more broadly, we gap filled the wind speed and temperature data using a self-organizing linear output (SOLO) map (Hsu et al. 2002), a type of artificial neural network (ANN). The drivers used were the best correlated variables of the same type from nearby locations, and we were generally able to represent about 90% of the variance of the target variables with the SOLO for wind speed, and more than 95% for temperature. No precipitation data were gap filled.
We performed an objective spatial analysis of the precipitation amount from the gauge data over the domain 36.95°–35.45°S, 147.85°–148.90°E, with grid spacing of 0.05°. The analysis was performed using the NCAR Command Language obj_anal_ic() routine, which is essentially a Barnes (1964) successive correction analysis. This spatial interpolation technique is related to kriging and assumes that the underlying data can be represented by a Fourier series. The accuracy of the technique is determined by the uniformity and accuracy of the input data, as well as the choices of the radii of influence parameters, which determine the convergence of the scheme in the successive iterations.
We used radii of influence of 0.05° and 0.15°, or one and three grid cells, for the two iterations. This means that the value determined for any given grid cell was most strongly influenced by gauges within 0.05°, or about 5 km, and no value was calculated if there was no gauge within 0.15° (15 km). The objective analysis was performed for each 6-h interval in the analysis period and summed for each year to give the total precipitation amount for April–October.
3. Partitioning of precipitation types for regression analysis
There are no direct observations of precipitation type at Guthega Dam. Measurements of humidity only began toward the end of the analysis period, so the only measured variable that may be used to categorize the precipitation type is the ambient temperature. Yang and Ohata (2001) used daily mean air temperature thresholds to define the precipitation type, with criteria set at T < −2°C and T > 2°C for snow and rain, respectively, with mixed precipitation between these bounds.
We defined optimal temperature intervals (defined by threshold values, or breakpoints, and ) based on analysis of the CR relationship between the NOAH II in the full DFIR fence (the reference gauge; PrC) and the unfenced NOAH II (the comparison gauge; PrB). Ordinary least squares (OLS) linear regression was used to evaluate the impact of 6-h average WS on CR for intervals , , and . The breakpoints were varied between −3° and 6°C with a step size of 0.1°C, under the condition that (the special case , or only two intervals, was considered separately). The RMSE was calculated for each of the breakpoint pairs (Fig. 4), and the values that minimized the RMSE overall were = 0.4°C and = 1.8°C.
A minimum threshold of 4 mm of precipitation at the reference gauge was used for this analysis, but the sensitivity to this parameter was low, with only varying between 0.2° and 0.5°C over thresholds of 1–8 mm. We repeated the analysis using NOAH II in the half DFIR fence as the reference gauge and produced the same result. Using data from the gauge pair at the Kerries produced a wider interval for mixed-phase precipitation ( = 0.2°C and = 2.6°C), but the low range of wind speeds there made it harder to achieve good-quality fits for the analysis.
There is good physical justification for these breakpoint values. Michelson (2004) used a large body of wintertime synoptic observations to derive a hyperbolic relationship for the fraction of snow based on wet-bulb temperature. The function was defined for −0.21° < < 2.42°C and has upper and lower decile values of 0.48° and 1.72°C, respectively. The agreement of the breakpoint analysis with these empirical values is consistent with the hypothesis that the NOAH II gauges perform differently for snow, mixed precipitation, and rain.
4. Developing optimal CR relationships between gauge pairs
Catch ratio relationships between various gauge pairs at Guthega Dam were calculated through OLS regression for the snow and mixed precipitation temperature intervals defined above. As a basic check, we first verified that the catch ratio for rain was very close to unity for each gauge pair. The 6-h intervals were included only if at least 2.0 mm of precipitation for the reference gauge and valid temperature and wind speed data were available for the whole interval. Different thresholds were tested; 1.0 mm affected the fits because the observed CR values were constrained to simple fractions (e.g., ½) due to the NOAH II gauge resolution. A threshold of 3.0 mm (used in a number of previous studies) did not show any improvement in quality of fit over 2.0 mm.
The predictors considered were the 6-h mean temperature, mean wind speed, the square of mean wind speed, and the total precipitation amount in the comparison gauge. Previous studies, analyzing 24-h intervals, often included daily maximum and minimum temperatures, but our use of subdaily intervals to resolve the diurnal cycle made this unnecessary. The use of the precipitation amount to predict CR is also somewhat novel, but we found that it was statistically significant as well as independent of the other predictors. We found that the square root of precipitation amount was a more statistically significant predictor than the simple amount.
Optimal predictor combinations were developed by minimizing the Bayesian information criterion (BIC; Schwarz et al. 1978), which prevents overfitting by penalizing complex models. We required all of the predictors to be statistically significant (P < 0.05) with respect to CR. There were a very small number (n < 4) of outliers in the raw CR values used to perform the regression for snow (specifically, CR ≥ 1.0 for moderate wind speeds). The fits obtained by including these outliers were judged to be nonphysical for higher wind speeds (WS ≥ 8 m s−1). The outliers mostly vanished when the precipitation threshold was set to 5.0 mm, but in order to retain the sample size we required that CR < 1.0 for wind speeds greater than 4 m s−1.
The following sections discuss the specifics of each of the CR relationships developed for the various gauge pairs considered. All of the CR relationships referred to are presented in Table 2.
a. Unfenced and full DFIR NOAH II gauges
Based on the 6-h intervals where both gauges were working optimally and the full DFIR recorded more than 0.1 mm of precipitation, the unfenced NOAH II gauge caught, on average, about 48% of the snow recorded by the gauge in the full DFIR fence and 77% of the mixed precipitation. This amounted to a total deficit of 1404 mm over the analysis period.
Robust catch ratio relationships were found between the unfenced NOAH II and the NOAH II in the full DFIR. Only one outlier was removed from the snow dataset where CR ≥ 1.0 for WS ≥ 4 m s−1.
These relationships are shown numerically in Table 2 and graphically against wind speed (Figs. 5a, 6a) for some constant values of T (which characterize the temperature intervals), and Pr = 5.0 mm (which, together with the fixed values for T, explained the discrepancy between the constant terms above and the y intercepts in the figures). The quality of the fits obtained indicates that the model explains the variance in CR very well for snow, and moderately well for mixed precipitation. The quadratic fit for snow had a minimum value for WS = 11.9 m s−1, which is at the extreme of the range of the wind speeds observed, making it a physically valid model.
The implication of the large Pr0.5 term is clear: the catch ratio was generally higher for greater precipitation amounts, for both snow and mixed precipitation. The physical explanation for this effect is that higher precipitation rates are associated with larger snowflake sizes, and hence greater fall speeds, in accordance with the Gunn–Marshall (Gunn and Marshall 1958) distribution (analogous to the better-known Marshall–Palmer distribution for raindrop size; Marshall and Palmer 1948). For snow, the wind speed was the next most important factor, with Fig. 5a showing that the catch efficiency of the unfenced NOAH II gauge dropped to around 0.3 for wind speeds above 8 m s−1. The coefficient of temperature for the mixed CR was more than 4 times that for snow, supporting the hypothesis that temperature had a controlling influence on particle fall speeds in this temperature range.
The optimal CR relationships with Pr omitted are presented in Table 2 for this gauge pair, but not the subsequent ones. Including Pr0.5 increased the absolute fraction of the explained variance in CR by 13% for snow and by 19% for mixed precipitation. Furthermore, the changes in the other coefficients were small, which indicates that the effect of the precipitation was nearly independent of the other effects. Finally, the change in the constant term may be reconciled against the precipitation amount term. For snow, the difference is neutralized by a precipitation amount of 4.95 mm and for mixed precipitation the amount was 5.26 mm, which are close the value of 5.0 mm used for plotting purposes in the figures.
b. Unfenced heated tipping-bucket gauge and full DFIR NOAH II gauge
The heated tipping-bucket gauge recorded 55% and 85% of the NOAH II gauge in the full DFIR fence for snow and mixed precipitation, respectively. This amounted to a total deficit of 1533 mm compared to NOAH II in the DFIR over the analysis period. Statistically significant relationships for CR were found for Pr0.5 and WS for both snow and mixed precipitation, but T was significant for mixed precipitation only (see Table 2 and Figs. 5b, 6b).
At moderate wind speeds (WS ~ 5 m s−1), the CR values for snow and mixed precipitation (see Figs. 5b, 6b) were similar to the unfenced NOAH II gauge, but the spread in the values was much higher, and the fit provided by the regression was poorer. At high wind speeds, the relationships suggest that CR dropped to zero for snow, and this is indeed borne out by the observations: zero precipitation amounts were common for the tipping-bucket gauge at wind speeds greater than 7 m s−1, while the reference gauge recorded at least 2.0 mm.
c. Half DFIR and full DFIR NOAH II gauges
The raw values of precipitation for snow and mixed precipitation recorded by NOAH II in the half DFIR fence were relatively unbiased, with 90% and 92% for snow and mixed precipitation relative to the NOAH II in the full DFIR fence. This amounted to a total deficit of 329 mm over the analysis period. The CR relationships are presented in Table 2 and Figs. 5c and 6c.
As the sole predictor (apart from precipitation amount), WS was found to have a statistically significant relationship with the CR, but WS2 was a slightly better fit. When both were used, the WS term became insignificant, so it was omitted from the above relationships. Temperature was not statistically significant for mixed precipitation.
The negative coefficient in the WS2 term is at odds with the CR relationships defined between reference and unfenced gauges in previous studies (e.g., Goodison et al. 1997), where the turning point represents a flattening of the CR at high wind speeds. However, the relationships above describe the performance of different fence configurations, so it is not necessarily valid to compare them with the relationships found in previous studies.
This result is important because it highlights some vagaries of using an alternative reference standard. The negative coefficient of temperature implies that the gauge inside the half DFIR fence was relatively more efficient than that with the full DFIR fence for drier snow. For both snow and mixed precipitation, the CR was approximately unity for low to moderate wind speeds, but the efficiency of the half DFIR fence decreased rapidly for higher wind speeds. There was some indication that the relative performance may level off at CR ≃ 0.5 for the highest wind speeds, but there were not enough data available to establish this.
d. Unfenced and half DFIR NOAH II gauges
The unfenced NOAH II gauge recorded 55% and 85% of the NOAH II gauge in the half DFIR fence for snow and mixed precipitation, respectively. This amounted to a total deficit of 1106 mm over the analysis period. Two outliers were removed from the snow dataset where CR ≥ 1.0 for WS ≥ 4 m s−1 to obtain the CR relationships (see Table 2 and Figs. 5d, 6d).
For this model (hereafter GD-only), the Pr0.5 term also had quite a marked effect for this gauge pair: omitting Pr from the regression resulted in an explained variance of 0.44 for snow and 0.09 for mixed precipitation. The coefficient for WS2 in the relationship for snow was somewhat smaller than in the relationship between the unfenced/full DFIR fence configuration. This can be explained in terms of the CR relationships described above. The positive curvature of the CR relationship between the unfenced NOAH II gauge and NOAH II in the full DFIR fence was offset by the relatively strong negative curvature of the relationship between the two fenced gauges. The minimum value for the CR relationship for snow occurs at WS = 12.4 m s−1, which is at the extreme of the observed wind speeds.
For mixed precipitation, the CR was higher at moderate to high wind speeds than when compared to the reference gauge. This also reflected the choice of reference gauge: we saw above that the efficiency of the NOAH II gauge in the half DFIR fence decreased with wind speed for mixed precipitation as well.
e. Inclusion of unfenced and half DFIR NOAH II gauges from the Kerries
Cross validation of the model presented for unfenced and half DFIR NOAH II gauges at other sites with multiple gauges (section 5b) showed that conditions at Guthega Dam do not represent more sheltered locations in the Snowy Mountains. The Kerries, on the other hand, had a mean wind speed of only 1.4 m s−1 and a 95th percentile value of 3.5 m s−1 during snow or mixed precipitation. The unfenced NOAH II gauge recorded 92% of snow and 98% of mixed precipitation compared to the NOAH II gauge in the half DFIR during the overlap period. Taken alone, the data from the Kerries were unsuitable for developing a CR model because of the small range in wind speed, so we combined the data from both the Kerries and Guthega Dam to develop a model we hypothesize to be more generally applicable in the Snowy Mountains (see Table 2, Fig. 7).
This CR model (hereafter GD+TK) departed from GD-only by having a higher catch ratio for low wind speeds, driven by the cluster of observations for WS < 2.0 m s−1 at the Kerries for snow, and the influence of temperature was also diminished. For mixed precipitation, the temperature coefficient was decreased by about 14% compared to Guthega Dam alone, and the relationship to wind speed was slightly steeper, again because of the cluster of low wind speed observations from the Kerries.
5. Evaluation and application of the derived CR relationships
a. Does applying a gauge adjustment improve the RMSE in comparison to a reference gauge?
The CR relationships developed above permit the estimation of the precipitation amount that would have been measured by a better-shielded reference gauge (PrR), by the simple application of the following expression:
where PRm is the measured precipitation amount and CR is the empirical catch ratio. Ultimately, this adjustment should be done only if it reduces the RMSE or bias. Table 3 contains the RMSE and mean bias before and after adjustment for each of the gauge pairs considered in section 4, both with and without a precipitation term. In addition to the CR fits, we evaluated the performance of a bias adjustment, based on a linear regression of precipitation amount only, for each gauge pair.
The reduction in overall RMSE for the unfenced and full DFIR (unfenced and half DFIR) NOAH II gauge pair was about 52% (49%), but if Pr0.5 was omitted the reduction was only about 28% (26%). Including Pr0.5 in the CR relationship resulted in an undercorrection of 0.225 mm (0.150 mm), while omitting it resulted in an average overcorrection of 0.139 mm (0.123 mm).
For the unfenced and full DFIR gauge pair, both of the CR adjustments resulted in lower RMSE than the bias adjustment. For the unfenced and half DFIR pair, the CR adjustment omitting the precipitation term resulted in higher RMSE than the bias adjustment, but when the precipitation term was included the RMSE was lower than the bias adjustment.
For the heated tipping-bucket gauge, the bias adjustment was only narrowly beaten by the CR adjustment when Pr0.5 was included, and when it was omitted the overall RMSE was actually increased. For the half DFIR and full DFIR NOAH II comparison, the CR adjustments reduced the RMSE by about 15%, regardless of whether precipitation was included, whereas the bias adjustment had a smaller impact.
The mean RMSE of the GD+TK raw data was lower than for Guthega Dam alone, because the precipitation data from the unfenced NOAH II gauge at the Kerries were relatively accurate because of the low wind speeds there. The bias adjustment resulted in only a modest reduction in RMSE for the same reason. The CR adjustments, both with and without Pr0.5, resulted in an overall decrease in RMSE.
b. Cross validation of adjustment scheme
The empirical CR adjustment scheme presented above requires cross validation before it may be robustly applied to other sites. In this section, we concentrate on the unfenced and half DFIR fenced NOAH II gauge pairs, simply because there are no other full DFIR fences in the Snowy Mountains (or anywhere else in the Australian Alps), so it is not possible to validate the full DFIR model thoroughly. Furthermore, when we apply the adjustments more broadly in section 5c, the use of a half DFIR model allows us to homogenize the network with the smallest number of adjustments.
We directly cross validated two adjustments derived for half DFIR fenced versus unfenced NOAH II gauges (GD-only and GD+TK) using independent data from Grey Hill and Snowy Plain, with RMSE and bias of the adjusted precipitation amounts shown in Table 4. Both adjustments resulted in larger overall RMSE at the independent sites. The GD-only model was considerably worse at both sites, with moderate increases in RMSE for both snow and mixed precipitation. For the GD+TK model, the changes in RMSE for Snowy Plain were negligible. There was a modest decrease in RMSE for mixed precipitation at Grey Hill, but a larger increase for snow dominated the overall RMSE. The adjustment performed resulted in a large overcorrection, increasing the mean bias from −0.05 to 0.73 mm.
This result is not particularly encouraging, but there were some issues with the cross-validation sites that should be discussed. In particular, we believe that Grey Hill is a poor candidate for a cross validation site. It was the only site at which the CR for the unfenced NOAH II was higher for snow than for mixed precipitation. It is difficult to evaluate this anomaly because wind speed was only measured at the fenced gauge, but there were no obvious differences in terrain or vegetation that could account for influences on local wind flow. Additionally, both sites had relatively low wind, so the mean bias in the raw data was very low to begin with. It is unfortunate that more exposed, reliable cross-validation sites were not available.
To address the issue of suboptimal cross-validation sites, we also tested the models using a jackknife method (Tukey 1958) to develop statistically independent data from the intercomparison sites themselves. The method consisted of removing a single interval from the GD-only and GD+TK datasets and deriving the CR relationship with the remaining data and testing this against the withheld (independent) data. This was repeated for each interval in which any precipitation was reported, and RMSE and bias in Table 4 were calculated in the same way as previously.
We found that the validation of the GD+TK model with data from Guthega Dam resulted in only a very small increase in RMSE compared to the GD-only model, whereas the RMSE for the cross validation using data from the Kerries increased when using the GD-only model but decreased using the GD+TK model. When validated using the combined GD+TK dataset, the overall RMSE was lower for the GD+TK model. The model coefficients were stable with respect to the omission of single data intervals, and varied by less than 5% (95% confidence interval).
c. Application of the results to the Snowy Mountains network
The results of applying the GD+TK CR relationship to each of the 12 sites with an unfenced NOAH II gauge are shown in Table 5. The conditions at the sites are quite varied, ranging from the relatively warm and calm conditions of the Murray 1 Valve House to exposed sites on the Main Range of the Snowy Mountains. For all of the sites below 1400 m MSL in elevation (the approximate snow line), the estimated undercatch was 5% or less of the total precipitation amount. The mean temperature during precipitation at these sites was 4°–5°C, and snow and mixed precipitation accounted for around 15%–30% of the total measured precipitation.
For sites above 1400 m MSL, about half of the measured precipitation was snow or mixed (except at Guthega Dam, which is in a valley directly downwind from the Main Range and may be warmed by foehn winds when precipitation occurs along the Main Range). Although the Kerries was the equal coldest of the sites on average, it had the smallest adjustment of the high-altitude sites because of the light winds there. The adjustments had an impact of about 9% of the total recorded amounts at Pinnacle Mountain and Grey Hill, which were relatively warm and unexposed. Mount Hudson, on the other hand, is the second highest and the windiest site. Here, we estimate that an adjustment of about 52% of the total recorded wintertime precipitation is needed to represent the amount that would have been recorded by a NOAH II gauge shielded inside a half DFIR fence.
d. Spatial analysis of precipitation amount
An objective spatial analysis was performed on the precipitation gauge data from a total of 70 gauges at 61 unique sites in the Snowy Mountains region, using the method described in section 2. The gauges used included
22 unheated tipping-bucket gauges, with elevation up to 1250 m MSL;
15 heated tipping-bucket gauges, with elevation between 800 and 1320 m MSL;
12 unfenced NOAH II gauges; and
21 NOAH II gauges in a half DFIR fence.
The unfenced NOAH II gauge data were adjusted using the methods described above to estimate the precipitation that would have been recorded had a half DFIR fence been installed. The tipping-bucket gauge data were not adjusted, but the effects at lower altitudes were judged to be minimal (see Table 5). The resultant gridded dataset was designated the half DFIR analysis.
To test the overall effect of the wind fences installed in the Snowy Mountains region, the objective analysis procedure was repeated, but this time applying the inverse of the catch ratio calculated to the NOAH II gauges in half DFIR fences, and leaving the unfenced NOAH II gauges unadjusted, to create a no DFIR analysis. The two analyses are compared in Fig. 8.
Differences at low elevations were negligible because the tipping-bucket gauge data were identical in each analysis. Above 1000 m MSL, the mean difference between the analyses was 43 mm per season, or about 6% relative to the no DFIR analysis. The maximum difference between the two analyses was 249 mm, which is 25% relative to the no DFIR analysis at the same grid point.
a. Uncertainty due to precipitation phase classification scheme
There are several justifications for using a temperature-only criterion in this paper instead of a humidity-aware scheme (e.g., Michelson 2004; Harder and Pomeroy 2013). First, no humidity measurements were made at Guthega Dam during our analysis period, but a sensor was installed in 2013 during the reconfiguration of the site for the SPICE. For 224 six-hour intervals in the winter months of 2013/14 with at least 4 mm of precipitation (the threshold used in section 3), the wet-bulb and ambient temperature were very highly correlated (R = 0.99). This suggests that ambient temperature is indeed an adequate discriminant for precipitation phase, at least in an alpine environment during wintertime. However, we found that 2 years of data were insufficient to produce a robust result in repeating the analysis of section 3 with wet-bulb temperature. Second, since humidity was only measured at a selection of sites, most of the phase discrimination in a humidity-aware scheme would be based on estimates, which is another potential source of error. On the other hand, ambient temperature was directly measured at nearly every one of the sites considered in this paper.
Since (as for hydrometeor temperature), a temperature-only scheme could result in misclassification of snow as mixed precipitation, and of mixed precipitation as rain, but not the reverse. For a given temperature, the empirical CR for snow was less than for mixed precipitation, meaning that such a misclassification would always result in overestimate of CR, and an undercorrection would be applied. Using only ambient temperature is therefore a conservative approach and is more desirable given the assumptions that would need to be made in deriving a networkwide humidity estimate for the Snowy Mountains.
b. Uncertainty due to spatial interpolation of precipitation
Regardless of the scheme used, spatial interpolation of precipitation data can lead to large errors, especially in mountainous terrain. Aside from the condition of the input data, local topography can have a profound effect on precipitation amounts (as we have seen in this paper). Objective analysis of precipitation amount implicitly assumes that conditions in any grid cell are similar to those in the next, which is in general not the case, but on the scale of about 5 km this is probably not unreasonable for the Snowy Mountains region. In any case, the spatial interpolation scheme for the two analyses was consistent, so comparisons between the two were valid.
c. The impact of installing wind fences in the Snowy Mountains
Interestingly, there was a strong linear relationship between gridpoint elevation and the undercatch for grid points above 1000 m MSL (Fig. 9). We partitioned the gridded data into east and west of the continental divide (n = 85 and 115, respectively) and evaluated the undercatch for each group separately. The increase in undercatch with respect to height was similar for both groups, at 15.5 mm (100 m)−1 in the east and 17.4 mm (100 m)−1 in the west. The values were 0.41 and 0.50, respectively, implying high statistical significance given the number of samples. These relationships reflect the effect of both lower temperatures and increased exposure to winds experienced by the gauges at higher elevations.
The greatest precipitation amounts occurred along the Murray River catchment boundary (leftmost dark line in Fig. 8). This is also where the greatest absolute difference between the analyses occurs because of exposure to winds and a predominance of snow at high elevation. The secondary maximum for precipitation amount, located around 36.1°S, 142.3°E, is not associated with a large difference between the analyses. Despite the relatively high elevation of the sites in this region (1500–1700 m MSL), they are relatively protected, with mean wind speeds of 2–3 m s−1 when precipitation was recorded. There is a secondary maximum in the difference between the two analyses along the Snowy–Murrumbidgee catchment boundary (rightmost dark line in Fig. 8), where the sites in this region recorded mean wind speeds of 3–5 m s−1 during precipitation.
d. Comparisons with CR relationships from other studies
Goodison et al. (1997) provide specific CR relationships for a selection of different gauges compared to the WMO DFIR, which we compare to the relationships derived in this study in Fig. 10 (specifically, the unfenced NOAH II gauge compared to the NOAH II gauge in the full DFIR fence at Guthega Dam). For the purpose of plotting the values, we set Tmin = Tmax = Tmean = −5°C for snow, and 1°C for mixed precipitation. Our results assume the same nominal 5-mm precipitation amount as for previous figures.
We cannot exclude the possibility of minor differences in the performance of the NOAH II gauge in the full DFIR fence compared to the WMO standard Tretyakov DFIR, which have not been compared to our knowledge. In spite of this ambiguity, the CR for snow for the NOAH II gauge is within 5% of the values for the Tretyakov and NWS 8-in. gauges reported by Goodison et al. (1997) for wind speeds of 5.0 m s−1, which is close to the average at Guthega Dam during precipitation. Agreement with these two gauges was within about 10% for wind speeds up to about 7 m s−1. As values greater than this were absent from the analysis of Goodison et al. (1997), there is no value speculating on the divergence for higher wind speeds.
e. Comparisons with the networkwide impact in other studies
The no DFIR analysis presented in this paper was a theoretical exercise to determine the net effect of installing wind fences. We found that the magnitude of the losses of a network with no fences (6% on average over winter for elevations above 1000 m MSL, and 15% around the peaks) would be at the lower end of the scale of the results found in previous work by Yang and Ohata (2001), Zhang et al. (2004), Yang et al. (1998), Ye et al. (2004), and Yang et al. (2005), among others.
The results of this study are applicable to the Australian alpine region more generally. The Victorian Alps are the continuation of the Great Dividing Range to the south and are similarly important for hydrological inflows to the Murray–Darling basin. The Brindabella Range to the north of the Snowy Mountains, while considerably smaller, is the source of much of the inflows to the reservoirs of Canberra (the national capital). These mountains currently have very sparse tipping-bucket and manual gauge networks operated by the Australian Bureau of Meteorology (BoM). It would be possible to adjust these data in a similar way. However, we found that the RMSE for the tipping-bucket gauge adjustment was nearly twice that for the adjustment to the unfenced NOAH II gauges. Furthermore, the heated tipping-bucket gauge at Guthega Dam was a nonstandard installation, so further measurements would need to be obtained to characterize the BoM gauges. Installation of NOAH II gauges (or another standard similarly suited to snow measurement) at the existing BoM sites, and at a collection of new sites if possible, would permit the adjustment of historical data as well as improving the reliability of future precipitation measurements in Australia’s alpine regions.
In section 3 we showed that despite lacking humidity observations, we were able to partition the precipitation data into effective phases based on ambient temperature thresholds for the purpose of deriving the CR relationships. The temperature thresholds were chosen by minimizing the overall RMSE of a piecewise WS-only CR formulation. The analysis was robust to the choice of gauge pair at Guthega Dam, but returned a slightly different upper threshold for mixed precipitation at a second site, where the range of WS was low and the linear fits poorer. The thresholds derived by this novel technique are supported by empirical evidence from other studies, but an interesting follow-up study would be to evaluate this against humidity-aware schemes (e.g., Michelson 2004; Harder and Pomeroy 2013) when sufficient humidity data have been collected at Guthega Dam.
In section 4 we presented what we believe to be the best possible CR relationships between gauge pairs at Guthega Dam, over a wide range of meteorological conditions. We found it essential to include a term for precipitation amount to satisfactorily characterize the CR. This, to our knowledge, differs from previous studies, but there is a strong physical justification in the link between precipitation rate and particle size (Gunn and Marshall 1958; Marshall and Palmer 1948) and hence fall speed. Ultimately, the best model for the unfenced and half DFIR CR used data from a second intercomparison site (the Kerries) that together experienced a much wider range of wind speeds. We also showed that the half DFIR fence provides a similar level of protection from wind-induced losses to the WMO standard DFIR for the NOAH II gauges in low to moderate winds, but the performance deteriorates for higher winds.
We cross validated the unfenced and half DFIR model using independent data from other sites where a short overlap between fenced and unfenced gauges was available. This validation provided mixed results, with adjustments at one site (Grey Hill) moderately increasing the RMSE for snow, and the other (Snowy Plain) having little effect on RMSE but moderately reducing the mean bias. However, these sites were somewhat flawed in that they both experienced relatively low winds and thus low differences in raw catch amount, and the Grey Hill site was unusual in that it experienced a lower mean bias for snow than for mixed precipitation. In this sense, the neutral results for these sites probably represent a worst-case scenario for validation attempts.
A second cross-validation analysis used a statistically independent dataset derived from the cross-validation sites through jackknifing. Adjustments to this dataset resulted in substantial reduction in RMSE. We believe that, taken together, these cross-validation analyses support the extension of the CR adjustment scheme to the wider Snowy Mountains network.
We applied the CR relationships derived at the intercomparison sites to all of the unfenced gauge data to estimate the undercatch of individual gauges in a range of conditions in Table 5 and used the relationships to homogenize the NOAH II gauge network to account for changes in fence installations over 2007–12. To determine the overall impact of having the half DFIR fences installed, we also inverted the CR relationship to estimate what would have been recorded if no fences had been installed. The effects were as expected, with the biggest discrepancies of about 15% around the peaks, where frozen precipitation and high winds were most common, and discrepancies of 6% on average for regions above 1000 m MSL.
The authors would like to acknowledge funding from an ARC linkage grant (LP120100115) and assistance from Dr. Peter Isaac in implementation of the artificial neural network used to gap fill the surface meteorological data. This paper has benefited substantially from the insightful comments of two anonymous reviewers.