Computational fluid dynamics (CFD) studies have been used to simulate the airflow around precipitation gauges and the associated collection efficiencies for rain and solid precipitation (Nešpor and Sevruk, 1999; Constantinescu et al., 2007; Colli, 2014; Colli et al., 2014, 2015, 2016a, b; Thériault et al., 2012, 2015; Baghapour and Sullivan, 2017; Baghapour et al., 2017). These studies have demonstrated the influence of wind speed, turbulence, hydrometeor characteristics (size, density, drag, terminal velocity), and gauge and shield geometry on precipitation gauge undercatch. For rainfall, Nešpor and Sevruk (1999) showed increases in wind-induced error for smaller drop sizes with lower terminal velocities, with errors increasing for higher wind speeds. The conversion factor (inverse of integral collection efficiency) varied with the precipitation intensity and rainfall type, which influenced the distribution of hydrometeor sizes and terminal velocities. Thériault et al. (2012) demonstrated similar trends for snowfall, with collection efficiencies varying significantly with the type of solid precipitation and size distribution. Simulated collection efficiencies for wet snow and dry snow hydrometeors captured the general upper and lower bounds of experimental observations, respectively, with the lower collection efficiency for dry snow hydrometeors attributed to their lower terminal velocity and interaction with the local airflow around the gauge.

For a Geonor gauge with a single-Alter shield, Thériault et al. (2012) used a constant drag coefficient hydrometeor tracking model to develop a series of transfer functions based on wind speed for different hydrometeor types. Colli et al. (2015) extended this work to show the influence of different hydrometeor drag models on collection efficiency results. Empirical drag model results (Khvorostyanov and Curry, 2005), based on the relative hydrometeor-to-air velocity over the hydrometeor trajectory, were shown to yield higher collection efficiencies compared with constant drag coefficient results that can overestimate drag values. Colli et al. (2015) developed transfer functions based on wind speed for unshielded and single-Alter-shielded gauges for three specific hydrometeor size distributions. Further studies, using computationally intensive large eddy simulation (LES) models, better resolved the intensity and spatial extent of turbulence around the gauge orifice, which can lead to temporal variations in collection efficiency results (Colli et al., 2016a, b; Baghapour and Sullivan, 2017; Baghapour et al., 2017). The degree of turbulence was found to vary depending on the specific shield configuration and wind speed (Baghapour et al., 2017).

Intercomparisons of precipitation gauges have served as the primary mechanism for developing transfer functions. In the 1998 World Meteorological Organization (WMO) Solid Precipitation Measurement Intercomparison, transfer functions were determined experimentally by comparing measurements from different gauges (primarily manual) with those from a manual collector with a Tretyakov shield in the WMO Double Fence Intercomparison Reference (DFIR) configuration (Goodison et al., 1998). Precipitation events were monitored by observers, who reported the amount and type of snow, wind speed, and temperature statistics for each event. Events were defined based on the duration of continuous snowfall when the reference DFIR precipitation accumulation was greater than or equal to 3 mm. Adjustment functions for unshielded gauge collection efficiencies were recommended for snow, mixed precipitation, and rain, based on the wind speed at gauge height (Goodison, 1978; Goodison et al., 1998; Yang et al., 1998). While these adjustments could be applied to manual precipitation accumulation measurements, their application to automated measurements at shorter timescales, and where the precipitation type may not be well defined, presents a significant challenge (Colli, 2014; Colli et al., 2014, 2016a, b; Thériault et al., 2015, 2012).

The WMO commissioned another intercomparison, the Solid Precipitation Intercomparison Experiment (SPICE), to assess various automated technologies for the measurement of precipitation accumulation and snow depth and to recommend automated field reference systems (Nitu et al., 2018). An automated precipitation gauge configured with a single-Alter shield within a DFIR fence was chosen as the field reference configuration for precipitation accumulation; this was referred to as the Double Fence Automated Reference (DFAR) configuration. Transfer functions for unshielded and shielded gauges were derived as an exponential function of wind speed following the approach of Goodison (1978) and using 30 min precipitation events from the SPICE dataset (Kochendorfer et al., 2017a). Separate functions were developed for solid precipitation and mixed precipitation, as defined by air temperature ranges: less than -2 ∘C for solid precipitation and between -2 and 2 ∘C for mixed precipitation.

Using Bayesian analysis of Norwegian measurement data, Wolff et al. (2015) developed a precipitation phase-independent, continuous transfer function with respect to wind speed and air temperature for a single-Alter-shielded Geonor precipitation gauge. A similar, but less complex, function was developed by Kochendorfer et al. (2017a, 2018) using the SPICE dataset, including results from eight measurement sites in Canada, Norway, Finland, Spain, Switzerland, and the USA. The application of this “universal” function to precipitation accumulation measurements from unshielded weighing gauges in SPICE was shown to reduce the overall bias relative to the DFAR; however, reductions in the root mean square error (RMSE) were less significant (Kochendorfer et al., 2017a, b, 2018; Wolff et al., 2015).

When applying universal adjustments with wind speed and air temperature dependence, the errors can vary significantly by site, presumably driven by differences in climatology (Smith et al., 2020; Kochendorfer et al., 2017a). This has motivated further work on climate-specific transfer functions (Koltzow et al., 2020; Smith et al., 2020). Other studies have proposed the use of precipitation intensity for the improved adjustment of solid precipitation (Chubb et al., 2015; Colli et al., 2020). Another potential avenue for reducing errors in adjusted measurements is by improving the ability of transfer functions to distinguish among different precipitation types and their aerodynamic properties (Thériault et al., 2012; Wolff et al., 2015; Nešpor and Sevruk, 1999).

In this work, adjustment functions incorporating hydrometeor fall velocity are developed to reduce the uncertainty (RMSE) in collection efficiency and precipitation accumulation estimates from unshielded Geonor T-200B3 precipitation gauges. The unshielded gauge configuration allows for the assessment of a broader range of collection efficiencies, as the degree of undercatch is generally more pronounced for unshielded gauges relative to shielded configurations. Further, by focussing on the unshielded configuration, no assumptions are required regarding the behaviour of the shield slats and their role in momentum reduction and turbulence generation around the gauge.

A combined modelling and experimental approach is used in this study. In the modelling component, computational fluid dynamics and Lagrangian analysis is used to characterize the gauge collection efficiency dependence explicitly in terms of wind speed and hydrometeor fall velocity and to derive a corresponding transfer function. Details of the modelling work are included in the Supplement. In the experimental component, fall velocity and precipitation type estimates from the Precipitation Occurrence Sensor System (POSS) are used to investigate how the hydrometeor properties influence the relationships among measured catch efficiency, wind speed, and temperature. Two additional transfer functions are derived experimentally with wind speed and fall velocity dependence. These new transfer functions are assessed against transfer functions with dependence on wind speed and air temperature, including one of the universal functions developed by Kochendorfer et al. (2017a) and a climate-specific function derived herein using a similar methodology.

A computational fluid dynamics model was used to characterize the collection efficiency dependence with wind speed and hydrometeor fall velocity. The model is detailed in the Supplement (Sect. S1.1). Briefly, a high-resolution three-dimensional computer aided design model of the Geonor T-200B3 600 mm capacity gauge (hereafter Geonor gauge) with 2 m gauge orifice height was developed for the analysis. Time-averaged Navier–Stokes equations and a k–ε turbulence model with 5 % turbulence intensity at the inlet (Kato and Launder, 1993) were used to model the airflow around the gauge for horizontal wind speeds (Uw) from 0 to 10 m s-1, applied in 1 m s-1 increments. Separate simulations were conducted for each wind speed using monodispersed hydrometeors (Sect. S1.2) and size distributions for specified hydrometeor types (Sect. S1.3).

Experimental measurements were performed in conjunction with SPICE over the 2013/14 and 2014/15 winter periods (1 November to 30 April) at the Centre for Atmospheric Research Experiments (CARE) site in Egbert, Ontario, Canada. Measurements of precipitation accumulation were performed using 600 mm capacity Geonor T-200B3 gauges in unshielded and reference DFAR configurations. Both gauges were securely mounted on concrete foundations to limit wind-induced vibrations. The performance of these gauges was confirmed by full-scale field verifications at the start and end of testing, with annual maintenance to inspect, clean, level, and recharge each gauge. The gauges were charged with a mixture of antifreeze (60 % methanol and 40 % propylene glycol) and oil (Esso Bayol 35 in 2013/14, discontinued; Exxon Mobil Isopar M in 2014/15).

Measurements of precipitation occurrence were obtained using the Thies Laser Precipitation Monitor (LPM) installed inside the inner fence of the DFAR. Wind speed and direction measurements at 2 m gauge height were performed with a Vaisala WS425 ultrasonic wind sensor adjacent to the unshielded gauge. Temperature was measured with a Yellow Springs International model 44212 thermistor in an aspirated Stevenson screen. Further details are available in the SPICE final report (Nitu et al., 2018).

The instruments were sampled using a Campbell Scientific CR3000 data logger. For each Geonor T-200B3 precipitation gauge, the frequency and precipitation accumulation for each of the three transducers was reported at 6 s intervals, the latter computed from the former using manufacturer-provided calibration coefficients. Minutely measurements of precipitation occurrence from the Thies LPM were recorded. The scalar average wind speed and vector average wind direction were recorded over 1 min intervals. Based on SPICE procedures, these data were processed using a format check to replace missing data with null values, a range check to identify and remove outliers outside the manufacturer-specified output thresholds, a jump filter to remove spikes exceeding maximum point-to-point variation thresholds, and a Gaussian filter to smooth out high-frequency noise in Geonor precipitation accumulation measurements (Nitu et al., 2018). Periods of instrument maintenance and power outages were removed from the analysis. The Geonor accumulation data were aggregated to 1 min intervals for subsequent analysis.

Precipitation events were identified during both measurement periods using the SPICE event selection procedure (Nitu et al., 2018). These events were defined as 30 min periods with at least 0.25 mm of precipitation recorded by the reference DFAR precipitation gauge and at least 60 % precipitation occurrence reported by the Thies LPM. The use of the LPM as a secondary confirmation of precipitation occurrence minimizes the likelihood of events with false precipitation due to dumps of snow or ice into the gauge, wind-induced vibrations, or other factors. Following the approach of Kochendorfer et al. (2018), a minimum 0.075 mm accumulation threshold was applied for the unshielded gauge to ensure that measurements exceeded the gauge uncertainty and that derived collection efficiency values were reliable. The 30 min event duration was chosen to be sufficiently long to reduce noise and ensure high confidence in measured parameters and sufficiently short to avoid the influence of diurnal temperature variations, while also providing a larger number of events for analysis relative to longer durations. Note that unless otherwise stated, all precipitation events referred to hereafter are 30 min events derived using the above approach.

The POSS is a small upward-facing bistatic X band radar capable of measuring the precipitation fall velocity based on the Doppler frequency shift of the received signal (Canada, 1995; Sheppard, 1990, 2007; Sheppard et al., 1995; Sheppard and Joe, 1994, 2000, 2008). During periods of precipitation, the POSS outputs both the mean and mode received signal frequency derived from the Doppler frequency spectrum over the previous minute. The mean precipitation fall velocity (Uf_mean) is estimated from the transmitted wavelength (λ) and the mean frequency (fmean) of the measured Doppler power density spectrum for falling precipitation hydrometeors. 1aUf_mean=fmeanλ2 The mode precipitation fall velocity (Uf_mode) is described by a similar function, based on the mode frequency (fmode) of the measured Doppler power density spectrum. 1bUf_mode=fmodeλ2 For each 30 min event, the mean and mode event fall velocity correspond to the average of all minutely mean and mode values, respectively. The transfer functions presented in this work were derived using both forms of event fall velocity and assessed in terms of the RMSE and bias error (BE) of adjusted measurements relative to the DFAR. The specific fall velocity indicated for each transfer function corresponds to that which produced the lowest RMSE and BE. The POSS also provides a minutely precipitation type output corresponding to very light, light, moderate, and heavy precipitation for rain, snow, hail, and undefined precipitation. Each event is classified as “rain” or “snow”, corresponding to a minimum 70 % occurrence of that precipitation type over the event period (i.e. at least 21 min of precipitation occurrence). “Mixed” precipitation events correspond to the presence of both “rain” and “snow” for the remaining events not classified as rain or snow. “Undefined” precipitation corresponds to events where the precipitation is not captured by the three other classifications.

Due to the systematic error associated with gauge undercatch, the unshielded gauge can capture less precipitation than the true amount falling in the air. The measured collection efficiency (CEm) is defined as the ratio of the precipitation accumulation reported by the unshielded gauge (Pun) relative to that reported by the DFAR (PDFAR) for each event and is given by 2CEm=PunPDFAR. Assuming that the gauge measurement uncertainties are independent and random with equivalent accumulations (corresponding to a collection efficiency equal to 1) and uncertainties, the uncertainty in the collection efficiency (σCE) scales with the relative magnitude of the gauge uncertainty (σP) and the event accumulation value (P) by error propagation. 3σCE=2σPP Collection efficiency transfer functions attempt to capture the performance of the unshielded gauge relative to the reference configuration based on wind speed, temperature, or other meteorological parameters. They can then be applied to adjust precipitation accumulations from an unshielded gauge in operational settings where reference measurements are not available. 4Padj=PunCE Kochendorfer et al. (2017a, 2018) used SPICE measurement data from eight test sites to develop an exponential and trigonometric transfer function based on wind speed (Uw) and air temperature (T). This is referred to as KUniversal in this work (Eq. 5a). For wind speeds above a threshold value (Uwt) of 7.2 m s-1, the wind speed is fixed at the threshold value (Eq. 5b) to avoid the potential for erroneous catch efficiency values at higher wind speeds that were not well represented in the SPICE measurement dataset. Based on a similar rationale, no adjustment is applied for temperatures above 5 ∘C. Note that while Kochendorfer et al. (2017b) considered wind speeds at both gauge height and at 10 m, Uw will denote the gauge height wind speed in this work. 5aCEKUw≤Uwt,T=exp⁡-b1Uw1-tan⁡-1b2T+b35bCEKUw>Uwt,T=exp⁡-b1Uwt1-tan⁡-1b2T+b3 The coefficients for KUniversal are provided in Table 1.

Using the same formulation, a site-specific transfer function based on wind speed and temperature was derived using the CARE dataset, for comparison with KUniversal. Best-fit regression coefficients were determined by varying the temperature threshold below 5 ∘C with the collection efficiency constrained to 1 above the threshold value. Solving Eq. (5a) for the temperature when the collection efficiency equals 1 provides an additional constraint on the b3 coefficient as a function of the b2 coefficient and temperature threshold (Tt). 5cb3=tan⁡-1b2Tt-1 The coefficients for the CARE site-specific transfer function, referred to as KCARE in this work, are provided in Table 1. The temperature threshold was varied over the measurement range in 0.01 ∘C increments to provide the lowest overall RMSE.

Using the minutely POSS precipitation type output, events were classified as “rain”, “snow”, “mixed”, or “undefined” following the methodology in Sect. 2.4. The relative occurrence of different precipitation types as reported by the POSS for the event dataset is summarized in Table 2. The fall velocities in Table 2 were estimated by the POSS following the methodology in Sect. 2.4; the temperatures were estimated from a YSI44212 thermistor in an aspirated Stevenson screen as described in Sect. 2.2.

Based on the mean fall velocities and temperatures for each precipitation event (Fig. 1, Table 2), snow events occurred at temperatures below 0.5 ∘C and with fall velocities of 0.93 to 2.32 m s-1. Mixed events were characterized by mean temperatures between -7.0 and 2.1 ∘C and mean fall velocities between 1.2 and 4.6 m s-1, while undefined precipitation events occurred at mean temperatures between -5.4 and 6.6 ∘C and fall velocities between 1.0 and 4.3 m s-1. Rain events were characterized by mean temperatures between -4.8 and 18.9 ∘C and mean fall velocities between 1.4 and 6.4 m s-1. Over the temperature range between -5 and 2 ∘C, rain, snow, mixed, and undefined precipitation types were all present, demonstrating the challenge of estimating precipitation type using temperature alone (e.g. as done for the KUniversal and KCARE transfer functions). Within this temperature range, a wide variety of mean fall velocities, between 1 and 6 m s-1, is also apparent.

The numerical model results for monodispersed hydrometeors capture the three-dimensional airflow and hydrometeor kinematics and illustrate the reductions in collection efficiency with increasing wind speed and decreasing hydrometeor fall velocity (Fig. 2). These results demonstrate that collection efficiencies are similar for different hydrometeor types with different sizes, densities, masses, and drag values (spherical drag model) but similar fall velocities. This enables the characterization of collection efficiency independent of hydrometeor characteristics other than fall velocity, allowing for the broad application of transfer functions with wind speed and fall velocity dependence to various hydrometeor types.

A slight nonlinearity in the collection efficiency relationship with wind speed is apparent in Fig. 2, with the collection efficiency decreasing more rapidly at lower wind speeds and more gradually at higher wind speeds. This wind speed dependence has been demonstrated in previous studies (Nešpor and Sevruk, 1999; Thériault et al., 2012; Colli et al., 2016a; Baghapour et al., 2017) and is generally attributed to the three-dimensional velocity profile around the gauge influencing the trajectories and catchment of incoming hydrometeors. A strong nonlinear dependence on the hydrometeor fall velocity is apparent in Figs. 3 and 5. Hydrometeors with fall velocities above 5 m s-1 exhibit collection efficiencies close to 1, while lower hydrometeor fall velocities influence the rate of decrease of collection efficiency with wind speed. Collection efficiency decreases are most pronounced below 2.0 m s-1 hydrometeor fall velocity, where a wide range of collection efficiencies are possible. This demonstrates the challenge in adjusting liquid, solid, and mixed precipitation accumulations in situations where different hydrometeor types and sizes – and with very different fall velocities – can occur. These findings support the conclusions of Thériault et al. (2012), who demonstrated large collection efficiency differences across dry snow and wet snow hydrometeors with different terminal velocities. The present findings also support those of Nešpor and Sevruk (1999), who showed that the wind-induced error increases rapidly for smaller raindrop sizes with lower terminal velocities.

The CFD transfer function presented in Eq. (6) (coefficients in Table 1) is based on the computational fluid dynamics results for an unshielded Geonor T-200B3 600 mm capacity precipitation gauge for wind speeds up to 10 m s-1. The CFD transfer function captures the nonlinear change in collection efficiency well, with wind speed and hydrometeor fall velocity observed in the numerical model results across rain, ice pellet, wet snow, and dry snow hydrometeor types (Fig. 2). This expression was derived from simulation results up to 10 m s-1 wind speed and should be used with caution at higher wind speeds. Further, this transfer function has not been assessed experimentally for snow above 6 m s-1 wind speed in the present study for the CARE dataset. Adjusted precipitation accumulation estimates in this regime, where fall velocities are low and wind speeds are high, can be highly uncertain and should be treated with caution (Smith et al., 2020). Assessment of the transfer function at other sites under such conditions is an area for future work. Application to other gauge or shield combinations should also be investigated, as the flow dynamics around the gauge orifice are dependent on the specific gauge and shield geometry.

The CFD transfer function formulation based on the fall velocity can be applied broadly across rain and snow types for the unshielded Geonor gauge configuration. These results are based on time-averaged simulations, which provide an estimate of the mean velocities through the domain and have been shown to provide good overall agreement with experimental results (Baghapour et al., 2017). Further study using LES models, which can better resolve the eddy dynamics and temporal variations in the flow, and under different boundary conditions and turbulence scales representing different site conditions is recommended to better understand the collection efficiency under conditions with high wind speeds and low hydrometeor fall velocities.

Integral collection efficiency differences across precipitation types are small when stratified by wind speed and hydrometeor fall velocity (Fig. 5). This results from the ability of the hydrometeor fall velocity to capture differences in the integral collection efficiency across different hydrometeor types and precipitation intensities. The small differences in collection efficiency across different hydrometeor types with the same fall velocity are attributed to the varying contribution from higher fall velocity hydrometeors, with collection efficiencies approaching 1, in the mass-weighted distribution of hydrometeor fall velocities. The results in Fig. 5 follow the general nonlinear profile of the CFD transfer function (Eq. 6, Fig. 4), with the hydrometeor fall velocity defining the integral collection efficiency magnitude for a given wind speed. Results for the same wind speed range and precipitation types that are stratified by wind speed and precipitation intensity, and by wind speed alone, are provided in Sect. S2.2 and discussed in Sect. S3.2; these results show much larger variability across hydrometeor types relative to those in Fig. 5.

Transfer functions were derived using accumulated precipitation amounts reported by automatic weighing precipitation gauges over 30 min periods. This approach is consistent with that used in SPICE (Nitu et al., 2018) and the related derivation of transfer functions (Kochendorfer et al., 2017a). While automatic precipitation gauges can report at a temporal resolution of 1 min, or even higher, the extension of the transfer function derivation and evaluation to other temporal periods, or different accumulation thresholds, is beyond the scope of this work.

The Kochendorfer et al. (2017a) universal transfer function with wind speed and air temperature dependence, KUniversal, was derived from measurements at eight SPICE sites in the interest of making the transfer function broadly applicable across different climates. This broad applicability is furthered by the widespread availability of air temperature and wind speed measurements at meteorological stations. Recent studies have demonstrated that the performance of KUniversal can vary substantially by site (Smith et al., 2020). Therefore, climate-specific KCARE transfer function coefficients were also derived for comparison in the present study.

The KCARE transfer function has a lower temperature threshold and exhibits larger increases in collection efficiency with increasing temperature relative to KUniversal (Fig. 6a). These differences improved the overall RMSE for KCARE by reducing the overadjustment of some rain and mixed precipitation events; however, this improvement came at the expense of underadjusting some snow events at warmer temperatures. The use of this approach warrants further study over longer periods to better understand the performance impacts of seasonal variability and assessment at other sites and climate regions with different precipitation characteristics and proportions.

Both the KUniversal and KCARE transfer functions performed well for snow but were limited by their ability to distinguish among snow, rain, and mixed precipitation at temperatures between -5 and 2 ∘C. The largest uncertainties in collection efficiency and adjusted accumulation estimates were observed over this temperature range. Adjustments using wind speed and hydrometeor fall velocity, however, addressed this shortcoming and provided improved collection efficiency and adjusted accumulation estimates. The CFD transfer function, derived from time-averaged numerical simulation results over a wide range of wind speeds and hydrometeor fall velocities, resulted in low RMSE values overall and across rain, snow, mixed, and undefined precipitation types. These results reinforce the fundamental importance of both wind speed and hydrometeor fall velocity on gauge collection efficiency demonstrated by the CFD model results and results from earlier studies (Nešpor and Sevruk, 1999; Thériault et al., 2012).

The CFD transfer function exhibited the lowest RMSE of all transfer functions for mixed precipitation and for intermediate fall velocities between 1.5 and 2.5 m s-1 (Table 4c), which is attributed to its nonlinear increase in collection efficiency with fall velocity. As this transfer function was derived theoretically, it is applicable across different sites and climate regimes with different types and relative proportions of hydrometeors. The present results also support the methodology for the CFD model, which can be extended to other shield and gauge combinations. For larger shields, it may be important to employ a more realistic vertical wind profile, with a zero-slip boundary condition at the earth's surface.

The HE1 transfer function showed good results for snow, supporting its use for the unshielded gauge. This approach is straightforward to implement based on its simplicity and is less reliant on the accuracy of fall velocity estimates beyond the fall velocity threshold. The collection efficiency for the HE1 transfer function decreases to 0.2 at a wind speed of 5.75 m s-1. This demonstrates the challenge of adjusting unshielded gauge snow measurements at windy sites, where the captured accumulations may be small relative to gauge uncertainties. This can lead to large uncertainty in adjusted measurements, as demonstrated by other studies applying transfer functions to unshielded gauge measurements at windy sites (Smith et al., 2020). The CFD transfer function results suggest a gradual decrease in collection efficiency at higher wind speeds compared with the HE1 transfer function, as some hydrometeors with higher fall velocities are still able to be captured by the gauge; however, these accumulations remain small relative to gauge uncertainties, particularly in windy conditions, making them difficult to assess experimentally. Further testing at other sites is recommended to better understand the collection efficiency for low fall velocity hydrometeors (light snow) under windy conditions above 6 m s-1, which were not available in the CARE dataset.

A limitation of the HE1 transfer function is the minimal improvement in the RMSE for mixed precipitation and fall velocities between 1.5 and 2.0 m s-1 relative to the KCARE function. This is due to the overadjustment of mixed precipitation events with fall velocities slightly below the cutoff value and the underadjustment of mixed precipitation events with fall velocities slightly above the cutoff. While the RMSE for mixed precipitation is still lower than that for adjustments based on temperature and wind speed (KUniversal, KCARE), further improvements are obtained using transfer functions with continuous fall velocity dependence – specifically, the CFD and HE2 transfer functions.

The HE2 transfer function, with a linear increase in collection efficiency with fall velocity, yields a greater reduction in the RMSE for mixed precipitation relative to the HE1 transfer function. The HE2 transfer function results show a higher RMSE for mixed precipitation than those for the CFD function, possibly due to the nonlinearity in the latter with fall velocity. The HE2 transfer function, however, yields the best RMSE results for snow, temperatures below -5 ∘C, and fall velocities below 1.5 m s-1. Adjusted uncertainties for snow are approximately 2 times higher than those for rain and show similar trends with increasing temperature and decreasing fall velocity. The former may be due to the lower event accumulations and greater adjustments for snow relative to rain, with measured values in closer proximity to the gauge uncertainty. The present approach of estimating the fall velocity using the POSS appears to perform well, overall; however, further study to better characterize the fall velocity distribution and changes over 30 min time periods could lead to further improvements in the model under specific conditions such as mixed precipitation. While this transfer function was derived using the CARE dataset, it is more universally applicable than adjustments based on temperature, for which the relative proportions of rain, snow, and mixed precipitation at warmer temperatures can influence fit results. Further testing at other sites is recommended to assess this in different climate regions, with different hydrometeor types and associated fall velocities.

It is evident that the performance of catchment-type precipitation gauges is dependent on wind speed and the aerodynamic properties of both the gauge and incident hydrometeors (Nešpor and Sevruk, 1999; Thériault et al., 2012; Colli et al., 2016b). The modelling results of this study demonstrated this dependence from a theoretical perspective, resulting in a transfer function that incorporates hydrometeor fall velocity. The experimental results validated this approach, which resulted in improved precipitation estimates from an unshielded gauge relative to those using surface temperature as a proxy for precipitation phase or type. Indeed, the use of surface temperature in this manner can be instructive (Kienzle, 2008; Harder and Pomeroy, 2013) but does not capture the conditions defining hydrometeor initiation and growth aloft (Stewart et al., 2015).

In this study, the fall velocity of hydrometeors reported by the POSS provided direct measurement of a key parameter related to the aerodynamics of the catchment process. In Canada, the POSS was deployed operationally to report present weather as part of an automatic weather station. In operational monitoring networks, the hydrometeor fall velocity can be provided by disdrometers (Loffler-Mang and Joss, 2000; Sheppard and Joe, 2000; Bloemink and Lanzinger, 2005; Nitu et al., 2018), vertically pointing Doppler radars (Biral, 2019), or multi-frequency radar techniques (Kneifel et al., 2015). Globally, other types of disdrometers (e.g. OTT Parsivel2, Thies Laser Precipitation Monitor) have been deployed operationally and can also provide hydrometeor fall velocities. The uncertainty in fall velocity estimates for different technologies, hydrometeor types, sizes, fall velocities, wind speeds, and wind directions remains to be assessed. These sensors can also be useful for reporting present weather and verifying the occurrence of precipitation based on their high sensitivity (Nitu et al., 2018; Sheppard and Joe, 2000).

The results from this study demonstrate that the combined use of accumulation reports from an unshielded weighing gauge with fall velocities reported by a disdrometer, wind speed measurements, and an appropriate transfer function can greatly reduce the uncertainty of precipitation accumulation measurements. The extension of the approach in the present study to shielded precipitation gauges or gauge designs with higher sensitivity may provide a means of further reducing the measurement uncertainty for automatic gauges in windy environments. Application to light snow events and different event durations are other areas for future study.