1. Introduction
Determination of winter precipitation type is a significant forecast challenge (Ralph et al. 2005), especially in thermal environments where rain (RA), snow (SN), freezing rain (FZRA), or ice pellets in the form of sleet (PL) all may be possible. Impacts to society and the economy from different precipitation types also can vary significantly. For example, a few millimeters of liquid-equivalent precipitation in the form of FZRA could produce much greater impacts than this same amount falling as RA, SN, or PL (Changnon 2003).
This challenge is compounded by increasing demands for high temporal and spatial resolution gridded forecasts. The NWS provides much of its forecast information through the National Digital Forecast Database (NDFD; Glahn and Ruth 2003). The NDFD offers forecasts of sensible weather elements (e.g., cloud cover, temperature, precipitation type) on a horizontal grid of 2.5-km resolution, often in hourly increments. The database is populated and regularly updated by meteorologists at local NWS offices around the country so that it reflects current observational trends as well as the most likely forecast scenario. The portion of the NDFD maintained by each office contains thousands of grid points.
At longer time ranges NWS meteorologists rely primarily on numerical weather prediction models to guide their forecast decisions. At close time ranges observational datasets play a larger role. In both cases, deviations from model guidance often are necessary, though interelement consistency within the NDFD must be maintained. For example, if hourly temperatures at certain grid points are adjusted from above freezing to below, the precipitation types and accumulations also must change for those hours and grid points. In fact, managing the hourly evolution of precipitation type across thousands of 2.5-km grids is one of the more challenging tasks an NWS forecaster faces.
The frequent need to adjust away from model guidance also is what motivates the use of implicit techniques based on vertical thermal profiles rather than explicit techniques utilizing model microphysics. Reeves et al. (2016) discusses a number of limitations related to model microphysics schemes. One not mentioned is that when a model is struggling to match surface and upper-air observational trends, explicit precipitation-type output based on its microphysics will misrepresent what is actually occurring at a particular time and location. Tools for making real-time forecast adjustments rely on the latest observations and do not have access to the internal microphysics of a model.
This paper describes a simple but effective implicit technique to assist in the operational forecasting of precipitation type. This is based on a revised version of the Bourgouin (2000) area method, referred to here as the Modified Bourgouin method. Section 2 describes three precipitation-type algorithms: the spectral bin classifier (Reeves et al. 2016), the traditional top-down approach used by many NWS offices (Baumgardt 2000), and the original Bourgouin area method. Section 3 explores motivations for a revised version of the Bourgouin approach. Section 4 describes the development of this revised algorithm. Section 5 shows how the revised technique performs relative to the other three. Section 6 provides a discussion and conclusions.
2. Precipitation-type algorithms for operational forecasts
a. Precipitation-type environments
Precipitation-type forecasts consider two key questions: whether there will be ice nucleation and whether there will be phase changes before hydrometeors reach the surface. The temperature profile in the lower troposphere can play a large role in determining the phase change potential (Czys et al. 1996) though the wet-bulb temperature Tw profile can be even more important (Ramer 1993; Baldwin et al. 1994; Schuur et al. 2012; Reeves et al. 2016) likely because Tw is more representative of the temperature experienced by particles large enough to survive evaporation.
Figure 1 summarizes four basic scenarios that can evolve when ice nucleation does occur. In the simplest case (Fig. 1a) the entire vertical profile is below freezing which supports SN. With a surface-based melting layer (layer with Tw above 0°C, Fig. 1b) the outcome is RA and/or SN depending on the depth and magnitude of that layer. An elevated melting layer (Fig. 1c) supports SN, FZRA, PL, or mixtures of all three. With another melting layer at the surface (Fig. 1d), RA replaces FZRA.
Wet-bulb temperature Tw profiles typically associated with (a) SN; (b) RA and/or SN; (c) FZRA and/or PL and/or SN; and (d) as in (c), but with RA instead of FZRA. Layers of melting energy (ME) and refreezing energy (RE) also are shown. Example heights along the y axis are representative values, but in the real world there are no limits on how these may vary.
Citation: Weather and Forecasting 36, 2; 10.1175/WAF-D-20-0118.1
In addition to the temperature or wet-bulb profile, hydrometeor sizes also can affect whether a phase change occurs (Crawford and Stewart 1995; Bernstein 2000; Cortinas 2000; Rauber et al. 2000, 2001; Robbins and Cortinas 2002; Changnon 2003; Cortinas et al. 2004; Thériault et al. 2010). In real time the hydrometeor size distribution is difficult to account for given the lack of sampling and thus is neglected in many precipitation type methods.
b. The spectral bin classifier
One algorithm that does account for hydrometeor size and has been shown to outperform other common implicit methods is the spectral bin classifier (SBC; Reeves et al. 2016). The SBC first considers the potential for ice nucleation then calculates melting and refreezing rates for hydrometeors of various sizes as they fall. It chooses one most likely precipitation type out of six possible categories: SN, RA, FZRA, PL, RASN, and FZRAPL. It does not provide information regarding the relative probability of alternate types.
While the SBC explicitly accounts for hydrometeor size during melting and refreezing, size is not considered during the initial determination of ice nucleation. Of note, the spectrum of hydrometeor sizes used by the algorithm is based on a static and theoretical distribution of drop sizes, not one determined uniquely for each situation. Additionally, the algorithm requires a detailed vertical temperature or wet-bulb profile. On the 2.5-km NDFD grid it would be difficult to adjust in real time a complete profile at each point. This algorithm therefore may be more suited to postprocessing of NWP output than manipulation by operational forecasters. Its skill relative to other approaches warrants inclusion in this discussion for comparison purposes.
c. The traditional top-down approach
A simpler approach to diagnosing phase change potential in a layer is to consider only one temperature variable: the maximum temperature or wet bulb in the melting layer, and the minimum temperature or wet bulb in the refreezing layer. This is the traditional top-down approach used in many NWS offices (Baumgardt 2000). As with the SBC, this also considers the potential for ice nucleation. Unlike the SBC it does not directly account for hydrometeor size. It does allow for certain combinations of mixed events by providing information regarding the relative probability of multiple precipitation types (RA, SN, FZRA, and PL) rather than trying to determine a single most likely outcome.
The traditional top-down approach works well for many scenarios. A key exception would be situations with isothermal layers in which the maximum or minimum temperature is less representative of the total melting or refreezing potential.
d. The Bourgouin method
Starting in 2016, some NWS offices have utilized an approach adapted from Bourgouin (2000). This improves over the traditional top-down method by accounting for both the depth and magnitude of a temperature layer, thus properly calculating the true melting or refreezing potential even in isothermal layers.
As with the traditional top-down approach, the Bourgouin technique requires a forecaster to manipulate only two variables (melting and refreezing energy) to adjust melting and refreezing potential. The advantage is that these variables more accurately reflect the phase change potential within a layer. And while the initial computation of energy values requires a detailed vertical profile, the use of simple layer-summary parameters facilitates subsequent real-time adjustments by the forecaster and calculations by the algorithm.
The next section describes the original Bourgouin approach in greater detail while also explaining several motivations for an enhanced version of this technique. Section 4 then describes the process utilized to develop this enhanced version.
3. Addressing limitations of the original Bourgouin area method
Since its introduction, the Bourgouin technique has been widely utilized by operational meteorologists via the BUFKIT program for examining forecast soundings (Mahoney and Niziol 1997) and in postprocessing of NWP output (Mullens and McPherson 2017; Wandishin et al. 2005). This technique is one of several algorithms Manikin (2005) combined into a single predominant precipitation-type value for use at NCEP (http://www.wpc.ncep.noaa.gov/wwd/impactgraphics/). Even so, the original version has a number of limitations that hinder its effectiveness. Some of these were discovered through the use of this tool for operational forecasts. Others were identified in a previous evaluation (Reeves et al. 2014) or anticipated in the paper which introduced this technique (Bourgouin 2000). The following discussion summarizes the original approach and explains motivations for a revised version of the algorithm.
a. The original Bourgouin area method
Using this approach, a sounding with no positive energy is classified as SN. If the only melting layer (defined as having at least 2 J kg−1 of positive energy) is surface based, the prediction is RA and/or SN depending on the magnitude of positive energy in that layer. With an elevated melting layer and surface-based refreezing layer the options are FZRA and/or PL depending on the ratio of melting and refreezing energy in the two layers. An additional surface melting layer would yield RA in place of FZRA.
b. Probability of mixed events
Like the SBC, the original Bourgouin approach chooses one most likely precipitation type out of six possible categories (SN, RA, FZRA, PL, RASN, and FZRAPL). Since information regarding the most likely alternatives would be of value operationally, the Modified Bourgouin approach always predicts the probability of all four basic types (RA, SN, FZRA, and PL). This allows for all varieties of wintry mixes (e.g., SNPL, FZRAPL, and FZRAPLSN) and gives information about the relative contribution from each basic type.
This probability of weather type (PoWT) approach also addresses other inherent uncertainties, either from NWP models or observed soundings. Even over a 2.5-km grid box and the span of 60 min, as forecasts are represented in the NDFD, there can be variations in the vertical thermal profile and thus in observed precipitation types. The PoWT approach offers insights regarding the most likely variations across space and time. In this way it also is more forgiving of observation or model errors, though like any technique will begin to fail as these errors become more egregious.
c. Developmental and independent datasets
The Bourgouin technique is a perfect-prog approach (Klein et al. 1959) derived from a database of surface observations and upper-air soundings. One factor acknowledged as potentially limiting the effectiveness of the original version is its relatively small developmental dataset. To establish thresholds separating FZRA from PL, the study utilized just 20 observations of PL, 31 of FZRA, and 3 of a mix. For distinguishing RA and SN the situation was somewhat better, with 53 observations of SN, 51 of RA, and 15 of a mix. The independent dataset for evaluation of the technique relative to other methods was even smaller.
To reassess these thresholds, a larger dataset is used. Our developmental dataset consists of 242 winter precipitation events across North America prior to 2006 (Table 1). To be used in the study, the reported precipitation event had to occur within a 35 km radius and within ±1 h of the time of a radiosonde observation, and surface observations had to be from locations with a human observer to ensure PL could be correctly detected (Landolt et al. 2019). Observations from Reeves et al. (2014) spanning the years 2002–05 were included in this dataset after an additional round of quality control. Many of the cases are from the mid-1980s through the 1990s, with a few dating back to the late 1970s.
Cases for each precipitation type in the present study.
An independent dataset for this new study is compiled from the years 2006–19. This includes the Reeves et al. (2014) events from 2006 to 2013. In total, there are 172 winter precipitation cases across North America in the new independent dataset (Table 1).
d. Temperature versus wet-bulb temperature
While the original Bourgouin method uses the vertical temperature profile for the calculation of layer energy, the revised version uses the Tw profile which likely offers a better discriminator (Ramer 1993; Baldwin et al. 1994; Schuur et al. 2012; Reeves et al. 2016) as previously discussed. Positive and negative Tw energy areas can be computed by substituting the environmental wet-bulb temperature (Tw-env) in place of Tenv in Eq. (1). This provides for the possibility of phase changes due to evaporative cooling in unsaturated layers.
e. Ice nucleation
The original Bourgouin method does not diagnose freezing drizzle, yet in Figs. 1a and 1c this or freezing rain is the most likely outcome if ice nucleation does not occur. Several studies have found that the activation of ice nuclei increases as cloud top temperatures drop through the range from −8° to −15°C (Rasmussen et al. 2002; Schichtel 1988; Politovich 1996; Pobanz et al. 1994; Bernstein 2000).
Probability of heterogeneous ice nucleation in a precipitation generation layer. Reproduced from Baumgardt et al. (2017).
Citation: Weather and Forecasting 36, 2; 10.1175/WAF-D-20-0118.1
f. Hydrometeor size
Unlike the SBC, neither the original nor the Modified Bourgouin approach directly considers hydrometeor size in any of its calculations. However, the revised technique does so indirectly by assuming there is a spectrum of hydrometeors in any winter precipitation scenario and that larger hydrometeors require more energy to prompt a complete phase change (Zerr 1997). Thus with increasing amounts of positive (negative) energy, the chance is greater that all hydrometeors will melt (refreeze).
4. Development of the Modified Bourgouin technique
The Modified Bourgouin technique as outlined above has several improvements over the original:
a larger developmental and independent dataset;
use of wet-bulb temperature profiles;
ability to diagnose freezing precipitation in situations without ice nucleation;
probabilistic output that accounts for uncertainties in the spectrum of hydrometeor sizes, forecast or observed soundings, and space and time, and
allowance for all combinations of wintry mixes.
a. Criteria for FZRA or RA and PL
A key limitation of the relatively small Bourgouin (2000) developmental dataset is the lack of PL and FZRA cases with large values of positive energy. This is especially true for PL events, with only one having more than 150 J kg−1. For this technique, the basic question is how much energy is required to produce a phase change. A related question for PL and FZRA is whether the magnitude of positive energy in the elevated melting layer influences the amount of negative energy needed to refreeze a hydrometeor. The original study suggests a completely melted hydrometeor requires considerably more negative energy to refreeze when there is a larger amount of positive energy. However, given the limited range of positive energy in the original dataset, this question is worth revisiting using a larger number of cases.
Figure 3 shows all cases in the new developmental dataset with FZRA and/or PL. Since the goal here is to determine thresholds between different precipitation types based off the Tw profile, only cases with at least an 80% chance of heterogeneous ice nucleation were included in the figure. This figure, similar to Fig. 2 from Bourgouin (2000), plots all events as a function of the positive/melting and negative/refreezing energies. However, energies here are based on Tw profiles. For reference the yellow line shows the original Bourgouin FZRA versus PL threshold.
Observations in the developmental dataset as a function of melting and refreezing wet-bulb energy. The yellow line is the Bourgouin (2000) function to separate FZRA from PL. The black line is the new threshold between pure FZRA and mixes containing PL.
Citation: Weather and Forecasting 36, 2; 10.1175/WAF-D-20-0118.1
Figure 4 shows distributions of elevated METw along with the surface layer RETw for FZRA, PL, and mixed events from the developmental dataset. While there is considerable overlap in METw magnitudes for these events, there is little to no overlap in RETw magnitudes between the two pure precipitation types. Pure FZRA events also have significantly less RETw than events containing any PL. Therefore, RETw is the most important predictor for the occurrence of PL once the hydrometeor is totally melted.
Wet-bulb melting and refreezing energy (ME and RE) for each precipitation type in the developmental dataset.
Citation: Weather and Forecasting 36, 2; 10.1175/WAF-D-20-0118.1
As in Fig. 3,. The shaded transition zones are used probabilistically to discriminate between FZRA and PL.
Citation: Weather and Forecasting 36, 2; 10.1175/WAF-D-20-0118.1
While the lower threshold [Eq. (6a)] is simply our derived logarithmic function from Eq. (5), the upper threshold defined here represents the point at which the chances of having an event of pure PL exceeds that of one mixed with FZRA. The RETw of 170 J kg−1 in Eq. (6c) represents the distribution crossing of the highest 30th percentile of mixed FZRA events and the lowest 30th percentile of pure PL events. Since this transition does not appear to follow a logarithmic function, we use a linear threshold with the same slope as defined in Eq. (3) above. This threshold also provided the best verification using the developmental dataset.
For FZRA, situations with barely any positive energy require special attention. In these cases PL and SN become increasingly favored, either because hydrometeors are only partially melted and thus easier to refreeze (Czys et al. 1996; Rauber et al. 2001; Reeves et al. 2016) or because hydrometeors experience little melting at all. To avoid over forecasting FZRA, the result from Eq. (7) is multiplied by 20% of the METw when melting energy is <5 J kg−1 (approximately the 10th percentile of METw for cases with FZRA). This is supported by the developmental dataset which shows the chances for cases involving FZRA falling at a rate near 20% for each 1 J kg−1 drop in the METw.
As a final step, the ProbIce adjustment is applied to results of Eqs. (7) and (8). ProbFZRA is adjusted by multiplying by ProbIce/100, then adding (100 − ProbIce). ProbPL is scaled downward via multiplication by ProbIce/100. The final probabilities again are constrained between 0 and 100.
The process of deriving probabilities for FZRA and PL from the melting and refreezing energies, along with the presence of ice, does result in a rather wide zone of possible mixed precipitation forecasts, as illustrated in Fig. 5. However, as previously mentioned, this is desirable since mixes of the two precipitation types are routinely observed in nature. The transition zone also captures the difference between pure FZRA and PL events.
b. Criteria for snow in wintry mixes
Most precipitation type algorithms do not attempt to identify the potential of SN in wintry mixes, especially in cases featuring an elevated warm layer. Recall that this is the case with the original Bourgouin technique, where only a forecast for FZRA, PL, or a mix of the two is possible when the elevated warm layer has more than 2 J kg−1 (Bourgouin 2000). Given that SN can be observed as part of a wintry mix, we seek to determine the METw magnitudes favorable for such. Figure 4 indicates that for all the mixed precipitation cases in which SN was observed (rightmost two types in Fig. 4), the median METw values were near or just below 10 J kg−1. This was in sharp contrast to cases void of SN, which were found to have significantly higher METw magnitudes. This suggests there is a threshold of METw that favors at least partial melting of all hydrometers, regardless of size.
Figures 6 and 7 display mixed precipitation cases with and without SN, respectively. These are similar to Fig. 5 but with the coordinates switched. In these figures, once METw exceeds 12 J kg−1 the number of cases with SN decreases significantly. In fact, 60% of the cases with SN occurred with METw at or below 12 J kg−1, and 75% occurred with METw at or below 20 J kg−1. This suggests that total melting of all hydrometers may not occur until METw exceeds 20 J kg−1, which is higher than identified in Bourgouin (2000).
Mixed precipitation cases including SN as a function of melting and refreezing wet-bulb energies (METw and RETw). The dark shaded blue color represents the threshold used to discriminate between wintry mixes containing SN vs those that did not.
Citation: Weather and Forecasting 36, 2; 10.1175/WAF-D-20-0118.1
As in Fig. 6, but for wintry mix cases not containing SN.
Citation: Weather and Forecasting 36, 2; 10.1175/WAF-D-20-0118.1
Interestingly, seven events had SN reported with METw in excess of 30 J kg−1. Four of these had METw over 89 J kg−1. It is unlikely that hydrometeors of any size would remain frozen with so much melting energy (Bourgouin 2000). We therefore explored the possibility this SN was generated in the cold surface layer and found the average RETw in these cases to be 173 J kg−1. This suggests these events had very cold and deep surface layers supportive of heterogeneous nucleation and SN generation in the cloud beneath the melting layer.
Given our goal is to identify a METw threshold to differentiate mixed precipitation events with SN from those without, the seven cases with surface-layer snow generation were removed from this analysis. Similarly, we removed the null snow cases that possessed METw in excess of 30 J kg−1. This left a total of 30 mixed events with SN and 54 without. Figure 8 provides METw distributions for these two samples. While the two distributions do overlap, a Mann–Whitney U-test revealed that these distributions were significantly different above a 99% confidence level. This provides evidence that only weaker METw magnitudes would support the presence of snow.
Wet-bulb melting energies (METw) for wintry mixes that contained SN and those that did not. Cases with METw above 30 J kg−1 are excluded.
Citation: Weather and Forecasting 36, 2; 10.1175/WAF-D-20-0118.1
While Eq. (9) was derived from cases involving a melting layer aloft, it can also be used to diagnose the potential for SN in environments with a melting layer near the surface. We tested Eq. (9) on 44 such cases and obtained a POD for SN of 0.86. This was in comparison to a POD of 0.57 using the original Bourgouin.
The appendix contains a summary of the steps involved in calculating probabilities for all four precipitation types.
5. Performance of the modified technique
This section evaluates the Modified technique’s performance using the independent dataset. We also compare its performance to that of the original Bourgouin area method, the SBC, and the traditional top-down methodology that utilizes MaxTwAloft. First we examine how well these algorithms predict FZRA and PL either alone or as part of a mix, and also how they predict SN as part of a mix. Then we see how closely their predictions match the full observations. For example, when only FZRA is reported we desire a forecast of FZRA alone, not a FZRAPL mix. Or when FZRAPL is reported we want only FZRAPL in the forecast.
This second, more stringent, evaluation is motivated by the way these forecasts can be utilized operationally, especially when QPF is derived independently of precipitation type. In that situation, the total QPF is distributed between different precipitation types according to the probability of each type. For example, if the algorithm yields a 100% chance of SN, all of the QPF would be assigned to SN. If, however, there was also a 30% chance of PL, then 77% (100/130) would go to SN and 23% (30/130) would go to PL. It is not desirable to split QPF in this way if only one type is actually occurring. The purpose of this more stringent test is to evaluate how well the algorithms predict all precipitation types in the observation.
Note that with this approach, a 50%/50% mix is equivalent to a 100%/100% mix. So there could be a forecast of 100% PL and 100% FZRA derived from Fig. 5. Or there could be a forecast of 100% PL that becomes a forecast of 50% PL and 50% FZRA after the ProbIce adjustment is applied. Both forecasts are equivalent for this evaluation.
The independent dataset for this study consists of 172 events from 2006 to 2019. Table 1 shows the number of cases for each precipitation type. While a larger dataset would lend greater confidence to these results, we seek a preliminary sense of how the different methodologies compare. This also is a far larger independent dataset than used in the original Bourgouin (2000) study.
For the first evaluation, the performance diagram in Fig. 9 allows for direct comparison of the different techniques by displaying four verification statistics: the probability of detection (POD), success ratio (SR, or 1 − false alarm ratio), bias, and critical success index (CSI) as described in Roebber (2009). The focus here is on any cases involving FZRA or PL either alone or as part of a mix, and SN as part of a mix.
Performance diagram for four precipitation-type methods: Modified Bourgouin (squares), original Bourgouin (triangles), MaxTwAloft (dashes), and spectral bin classifier (X marks). Forecast comparisons are for any FZRA (red), PL (purple), and SN (blue) in the observation. Curved lines represent the critical success index (CSI) while the diagonal lines represent bias.
Citation: Weather and Forecasting 36, 2; 10.1175/WAF-D-20-0118.1
Recall that the SBC and original Bourgouin methods are deterministic and provide a simple “yes” forecast of the one precipitation type or mix deemed most likely. In contrast, the traditional top-down and Modified Bourgouin techniques forecast the probability of four different types (RA, SN, PL, and FZRA) from which the probability of mixes also can be inferred. Consider a hypothetical scenario where the probabilistic output gives 60% for FZRA, 40% for PL and 30% for SN. If the actual observation is FZRAPL, then for Fig. 9 there would be a 0.6 hit for FZRA, a 0.4 hit for PL, and a 0.3 false alarm for SN.
In Fig. 9, a point nearer the center diagonal and upper-right corner shows less forecast bias (between POD and SR) and greater skill (CSI) than a point farther away. For FZRA, and for SN as part of a mix, the Modified Bourgouin technique produced the largest CSI of the four methods. Note that the SBC does not forecast SN when there is an elevated melting layer, and the MaxTwAloft method only does so when that layer is extremely small.
The clustering of PL scores at a lower CSI suggests this is a more difficult forecast for all the methods, and that all but the MaxTwAloft approach handle it similarly. Forecast skill for PL was very similar for the Modified Bourgouin technique and the SBC.
An important takeaway from Fig. 9 is the considerable improvement of the Modified Bourgouin method over the original. For FZRA the higher POD with little change in the bias is especially important given the considerable impacts associated with such events. This likely is due in part to the inclusion of a ProbIce check. In the independent dataset there were 31 FZRA cases with ProbIce ≤ 50. For these events the Modified Bourgouin POD was 0.96 while POD for the original was only 0.60. Similarly, when testing Eq. (7) on 60 FZDZ cases we obtained a POD of 0.92 with a CSI of 0.79. Moreover, the notable improvement to the Heidke skill score (HSS) for FZRA shown in Table 2 also indicates better forecasts of null events.
Heidke skill score (HSS) for any FZRA, PL, and SN in the observation from the traditional top-down max Tw, original Bourgouin, the Modified Bourgouin, and the Reeves SBC. The largest HSS for each type is shown in bold italics.
The diagram in Fig. 10 is similar to Fig. 9; however, the focus here is how well each algorithm exactly matches the observations. The performance for cases of FZRA and PL thus represent only those in which they were the sole precipitation type observed. Results for the various winter mixes involving the occurrences of FZRA and PL together (FZRAPL), and with SN (FZRAPLSN and PLSN), also are displayed.
As in Fig. 9, but for forecast comparisons for FZRA (red), PL (purple), FZRAPL (yellow), FZRAPLSN (light purple), and PLSN (light blue) for four precipitation-type methods: Modified Bourgouin (squares), original Bourgouin (triangles), MaxTwAloft (dashes), and spectral bin classifier (X marks).
Citation: Weather and Forecasting 36, 2; 10.1175/WAF-D-20-0118.1
Using Eqs. (10) and (11) with the same hypothetical scenario as before, the PureFZRA probability would be 33% [(60 − 40)/60], PureFZRAPL would be 17% [(40 − 30)/60], and PureFZRAPLSN would be 50% [(30 − 0)/60]. If the observation is FZRAPL then FZRA and PL are the desired forecast types. For purposes of Fig. 10 this yields a POD of 0.17, a false alarm of 0.33 for pure FZRA, and a false alarm of 0.5 for FZRAPLSN. If the observation is PL this forecast receives no credit in Fig. 10 since FZRA had a higher probability.
Results in Fig. 10 indicate that the Modified Bourgouin technique continues to display improved skill at predicting pure FZRA events over the original Bourgouin method. Also noteworthy is its superior skill at predicting wintry mixed events (FZRAPL, FZRAPLSN and PLSN), while having no adverse impact on the performance for pure FZRA events. Overall, with the exception of pure PL events, the Modified Bourgouin technique produced superior skill and HSS values for all precipitation types (Table 3).
Critical success index (CSI) and the Heidke skill score (HSS) for events of pure FZRA, pure PL, mixes of the two (FZRAPL), and for all combined precipitation types. The largest CSI and HSS for each type are shown in bold italics.
6. Summary and conclusions
The original Bourgouin technique (2000) is good at detecting FZRA events, but suffers from its inability to account for the potential lack of heterogeneous ice nucleation in the precipitation generation layer and evaporative cooling in unsaturated environments. The method also is hindered by its small developmental dataset. Therefore, in an effort to make the method a more viable precipitation type algorithm for operational forecasting, a modified version of the method was developed. This modified method is based on a much larger dataset, utilizes the wet-bulb profile instead of temperature, and incorporates a check for heterogeneous ice nucleation. New functions were empirically derived to discriminate between pure FZRA, wintry mixes, and pure PL. The modified method also allows for inclusion of SN despite an elevated melting layer. Together these changes yield significant performance improvements over the original Bourgouin and traditional top-down methods.
Because the Modified Bourgouin method was developed using observational data, it does not suffer from model bias and can be run with any numerical prediction model or with a blend of several models to produce gridded forecasts. While the revised method does not directly account for hydrometeor characteristics or the effects of precipitation rates, they are indirectly accounted for through the use of PoWTs. Preliminary verification results suggest its performance compares favorably to the more complex SBC that does attempt to account for hydrometeor characteristics. The Modified Bourgouin algorithm thus appears to show enough skill to be viable for NWP postprocessing, while also being simple enough to apply operationally for real-time forecast adjustments utilizing current observations.
An earlier version of the Modified Bourgouin technique was tested operationally at the NWS Weather Forecast Office (WFO) in Chicago for the winter of 2017–18, then as part of a larger testbed with over a dozen NWS WFOs during the 2018–19 and the 2019–20 winter seasons. Forecaster feedback has been positive. The earlier version also was adapted for versions 3.2 and 4.0 of the NWS National Blend of Models (NBM; Craven et al. 2018). The version of the Modified Bourgouin technique presented in this paper represents its latest iteration based on feedback from three anonymous reviewers and additional refinement by the authors.
Acknowledgments
The authors thank Dan Baumgardt, Andy Just, the NWS Central Region Forecast Builder team, and the NWS MDL for their assistance with this project, including adding the Modified Bourgouin algorithm into the NWS forecast tools (Forecast Builder) and adapting it for the NBM. We also thank Jeff Manion, NWS Central Region Headquarters, for a detailed review of the text. Three anonymous reviewers provided valuable comments and suggestions, which were greatly appreciated.
Data availability statement
All upper air and surface observations used during this study are openly available from the National Centers for Environmental Information at https://www.ncei.noaa.gov/.
APPENDIX
Calculating the Probability of Each Weather Type
a. Calculating ProbSN
Step 1:
. METw is all wet-bulb melting energy in the atmospheric column. ProbSNi is constrained between 0 and 100.
Step 2:
. ProbIce is the probability of ice presence in the precipitation generation layer. ProbSN is constrained between 0 and 100.
b. Calculating ProbPL (only when METw is nonzero)
Step 1:
. METw is the wet-bulb melting energy above a near-surface cold layer with nonzero RETw (wet-bulb refreezing energy). ProbPLi is constrained between 0 and 100.
Step 2:
. ProbIce is the probability of ice presence in the precipitation generation layer. ProbPL is constrained between 0 and 100.
c. Calculating ProbFZRA or ProbRA
Step 1:
. METw is the wet-bulb melting energy. RETw is the wet- bulb refreezing energy beneath any warm layers aloft. ProbFZRAi is constrained between 0 and 100.
Step 2:
. Step 3:
. ProbIce is the probability of ice presence in the precipitation generation layer. ProbFZRA constrained between 0 and 100. Where surface Tw > 0°C: ProbRA = ProbFZRA.
REFERENCES
Baldwin, M., R. Treadon, and S. Contorno, 1994: Precipitation type prediction using a decision tree approach with NMC’s mesoscale eta model. Preprints, 10th Conf. on Numerical Weather Prediction, Portland, OR, Amer. Meteor. Soc., 30–31.
Baumgardt, D., 2000: Precipitation type forecasting. Accessed 17 May 2020, http://rammb.cira.colostate.edu/training/visit/training_sessions/precipitation_type_forecasting/.
Baumgardt, D., A. Just, and P. E. Shafer, 2017: Deriving precipitation type probabilities in the national blend of models. 28th Conf. on Weather Analysis and Forecasting/24th Conf. on Numerical Weather Prediction, Seattle, WA, Amer. Meteor. Soc., 1190, https://ams.confex.com/ams/97Annual/webprogram/Paper313165.html.
Bernstein, B. C., 2000: Regional and local influences on freezing drizzle, freezing rain, and ice pellet events. Wea. Forecasting, 15, 485–508, https://doi.org/10.1175/1520-0434(2000)015<0485:RALIOF>2.0.CO;2.
Bourgouin, P., 2000: A method to determine precipitation types. Wea. Forecasting, 15, 583–592, https://doi.org/10.1175/1520-0434(2000)015<0583:AMTDPT>2.0.CO;2.
Changnon, S. A., 2003: Urban modification of freezing-rain events. J. Appl. Meteor., 42, 863–870, https://doi.org/10.1175/1520-0450(2003)042<0863:UMOFE>2.0.CO;2.
Cortinas, J. V., Jr., 2000: A climatology of freezing rain in the Great Lakes region of North America. Mon. Wea. Rev., 128, 3574–3588, https://doi.org/10.1175/1520-0493(2001)129<3574:ACOFRI>2.0.CO;2.
Cortinas, J. V., Jr., B. C. Bernstein, C. C. Robbins, and J. W. Strapp, 2004: An analysis of freezing rain, freezing drizzle, and ice pellets across the United States and Canada: 1976–1990. Wea. Forecasting, 19, 377–390, https://doi.org/10.1175/1520-0434(2004)019<0377:AAOFRF>2.0.CO;2.
Craven, J. P., and Coauthors, 2018: Overview of National Blend of Models version 3.1. Part I: Capabilities and an outlook for future upgrades. 25th Conf. on Probability and Statistics, Austin, TX, Amer. Meteor. Soc., 7.3, https://ams.confex.com/ams/98Annual/webprogram/Paper325347.html.
Crawford, R. W., and R. E. Stewart, 1995: Precipitation type characteristics at the surface in winter storms. Cold Reg. Sci. Technol., 23, 215–229, https://doi.org/10.1016/0165-232X(94)00014-O.
Czys, R., R. Scott, K. C. Tang, R. W. Przybylinski, and M. E. Sabones, 1996: A physically based, nondimensional parameter for discriminating between freezing rain and ice pellets. Wea. Forecasting, 11, 591–598, https://doi.org/10.1175/1520-0434(1996)011<0591:APBNPF>2.0.CO;2.
Glahn, H. R., and D. P. Ruth, 2003: The new digital forecast database of the National Weather Service. Bull. Amer. Meteor. Soc., 84, 195–202, https://doi.org/10.1175/BAMS-84-2-195.
Klein, W. H., B. M. Lewis, and I. Enger, 1959: Objective prediction of five day mean temperature during winter. J. Meteor., 16, 672–682, https://doi.org/10.1175/1520-0469(1959)016<0672:OPOFDM>2.0.CO;2.
Landolt, S. D., J. S. Lave, D. Jacobson, A. Gaydos, S. DiVito, and D. Porter, 2019: The impacts of automation on present weather-type observing capabilities across the conterminous United States. J. Appl. Meteor. Climatol., 58, 2699–2715, https://doi.org/10.1175/JAMC-D-19-0170.1.
Mahoney, E. A., and T. A. Niziol, 1997: BUFKIT: A software application toolkit for predicting lake-effect snow. Preprints, 13th Int. Conf. on Interactive Information and Processing Systems for Meteorology, Oceanography, and Hydrology, Long Beach, CA, Amer. Meteor. Soc., 388–391.
Manikin, G. S., 2005: An overview of precipitation type forecasting using NAM and SREF data. Preprints, 21st Conf. on Weather Analysis and Forecasting/17th Conf. on Numerical Weather Prediction, Washington, DC, Amer. Meteor. Soc., 8A.6, https://ams.confex.com/ams/pdfpapers/94838.pdf.
Mullens, E. D., and R. McPherson, 2017: A multialgorithm reanalysis-based freezing-precipitation dataset for climate studies in the south-central United States. J. Appl. Meteor. Climatol., 56, 495–517, https://doi.org/10.1175/JAMC-D-16-0180.1.
Pobanz, B. M., J. D. Marwitz, and M. K. Politovich, 1994: Conditions associated with large-drop regions. J. Appl. Meteor. Climatol., 33, 1366–1372, https://doi.org/10.1175/1520-0450(1994)033<1366:CAWLDR>2.0.CO;2.
Politovich, M., 1996: Response of a research aircraft to icing and evaluation of severity indices. J. Aircr., 33, 291–297, https://doi.org/10.2514/3.46936.
Ralph, F. M., and Coauthors, 2005: Improving short-term (0–48 h) cool-season quantitative precipitation forecasting: Recommendations from a USWRP workshop. Bull. Amer. Meteor. Soc., 86, 1619–1632, https://doi.org/10.1175/BAMS-86-11-1619.
Ramer, J., 1993: An empirical technique for diagnosing precipitation type from model output. Preprints, Fifth Int. Conf. on Aviation Weather Systems, Vienna, VA, Amer. Meteor. Soc., 227–230.
Rasmussen, R. M., I. Geresdi, G. Thompson, K. Manning, and E. Karplus, 2002: Freezing drizzle formation in stably stratified layer clouds: The role of radiative cooling of cloud droplets, cloud condensation nuclei, and ice initiation. J. Atmos. Sci., 59, 837–860, https://doi.org/10.1175/1520-0469(2002)059<0837:FDFISS>2.0.CO;2.
Rauber, R. M., L. S. Olthoff, and M. K. Ramamurthy, 2000: The relative importance of warm rain and melting processes in freezing precipitation events. J. Appl. Meteor., 39, 1185–1195, https://doi.org/10.1175/1520-0450(2000)039<1185:TRIOWR>2.0.CO;2.
Rauber, R. M., L. S. Olthoff, M. K. Ramamurthy, and K. E. Kunkel, 2001: Further investigation of a physically based, nondimensional parameter for discriminating between locations of freezing rain and ice pellets. Wea. Forecasting, 16, 185–191, https://doi.org/10.1175/1520-0434(2001)016<0185:FIOAPB>2.0.CO;2.
Reeves, H. D., K. L. Elmore, A. Ryzhkov, T. Schuur, and J. Krause, 2014: Sources of uncertainty in precipitation-type forecasting. Wea. Forecasting, 29, 936–953, https://doi.org/10.1175/WAF-D-14-00007.1.
Reeves, H. D., A. V. Ryzhkov, and J. Krause, 2016: Discrimination between winter precipitation types based on spectral-bin microphysical modeling. J. Appl. Meteor. Climatol., 55, 1747–1761, https://doi.org/10.1175/JAMC-D-16-0044.1.
Robbins, C. C., and J. V. Cortinas Jr., 2002: Local and synoptic environments associated with freezing rain in the contiguous United States. Wea. Forecasting, 17, 47–65, https://doi.org/10.1175/1520-0434(2002)017<0047:LASEAW>2.0.CO;2.
Roebber, P., 2009: Visualizing multiple measures of forecast quality. Wea. Forecasting, 24, 601–608, https://doi.org/10.1175/2008WAF2222159.1.
Schichtel, M., 1988: Specification of precipitation type in Oklahoma winter storms. M.S. thesis, Department of Meteorology, School of Meteorology, University of Oklahoma, 131 pp.
Schuur, T. J., H.-S. Park, A. V. Ryzhkov, and H. D. Reeves, 2012: Classification of precipitation types during transitional winter weather using the RUC model and polarimetric radar retrievals. J. Appl. Meteor. Climatol., 51, 763–779, https://doi.org/10.1175/JAMC-D-11-091.1.
Thériault, J. M., R. E. Stewart, and W. Henson, 2010: On the dependence of winter precipitation types and temperature, precipitation rate, and associated features. J. Appl. Meteor. Climatol., 49, 1429–1442, https://doi.org/10.1175/2010JAMC2321.1.
Wandishin, M. S., M. E. Baldwin, S. L. Mullen, and J. V. Cortinas Jr., 2005: Short-range ensemble forecasts of precipitation type. Wea. Forecasting, 20, 609–626, https://doi.org/10.1175/WAF871.1.
Zerr, R. J., 1997: Freezing rain: An observational and theoretical study. J. Appl. Meteor., 36, 1647–1661, https://doi.org/10.1175/1520-0450(1997)036<1647:FRAOAT>2.0.CO;2.
