Accurate precipitation measurements are crucial for a variety of applications, including water resource forecasting, future water availability, and hydrological and climate analysis and modelling (Barnett et al., 2005; Bartlett et al., 2006; Wolff et al., 2015). Canada's Changing Climate Report led by Environment and Climate Change Canada (Bush and Lemmen, 2019) highlights the importance of accurate precipitation measurements as fundamental climate quantities that play an important role in human and natural systems. Although the systematic bias due to the impact of wind on solid precipitation measurements is well documented (Goodison, 1978; Sevruk et al., 1991, 2009; Goodison et al., 1998; Yang et al., 2005; Smith, 2009; Wolff et al., 2015; Kochendorfer et al., 2017a), errors related to the automatic recording of precipitation measurements have only relatively recently been identified as automated weighing gauges become more commonly used (Sevruk and Chvíla, 2005). The cumulative precipitation data output from automated weighing gauges is subject to noise, diurnal temperature oscillations, and negative drift from evaporation, which can often mean that the precipitation signal over short sampling periods is influenced or hard to detect (Rasmussen et al., 2012). The nature of the noise and drift often varies substantially from site to site and between gauge configurations. High-frequency noise can exceed ±1 mm, and evaporation from the bucket can be in excess of several millimetres between precipitation events. It is therefore necessary to filter the raw data to separate real precipitation events from signal noise and identify and remove periods with evaporation (keeping in mind that evaporation reduces the precipitation amount derived from the differential in bucket weight). Improper filtering can lead to the accumulation of errors and result in significant inaccuracies in total seasonal precipitation. Duchon (2008) suggests that errors due to the diurnal oscillation in Geonor T-200B gauges could be 1 %–10 % of the precipitation total. Three post-processing challenges in the derivation of “clean” precipitation time series are the focus of this study: mechanical and electrical interference, diurnal oscillations, and evaporation of the bucket contents.

This study incorporates two commonly used accumulating automated weighing precipitation gauges (henceforth referred to as automated weighing gauges): the Geonor T-200B and OTT Pluvio2. The Geonor T-200B implements up to three vibrating wire transducers, which provide a frequency output that varies as a function of the fluid weight in the gauge bucket. The cumulative precipitation amount (bucket weight) is calculated from the frequency of each wire via calibration coefficients, with no onboard filtering (Geonor, 2019). The OTT Pluvio2 automated weighing gauge uses a high-precision load cell to weigh the bucket contents and provides several outputs, including intensity and precipitation accumulation (Nemeth, 2008; Nitu et al., 2018). The OTT Pluvio2 output has been pre-processed using an onboard proprietary algorithm which adjusts the high-frequency load cell measurements for temperature and vibration to derive a more accurate bucket weight. Further onboard processing removes the impact of unrealistic bucket-weight changes and evaporation from the output; however, this onboard algorithm was bypassed in this analysis to obtain the data in their rawest form.

A number of post-processing techniques have been developed to derive a noise-free precipitation time series from high-frequency automated weighing-gauge bucket-weight measurements. Some examples are described here.

The rolling maximum filter was used by Harder and Pomeroy (2013) to remove the “jitter” from the accumulated precipitation data sets by retaining a cumulative precipitation observation if it is greater than the previous maximum cumulative precipitation. The previous maximum is assumed to be the cumulative precipitation in all other cases. This filter reportedly works well in preserving the cumulative change in precipitation, but it may not always catch the precise start of precipitation events and will not always perform optimally in the presence of negative gauge drift (i.e. evaporation).

The World Meteorological Organization (WMO) Solid Precipitation Intercomparison Experiment (SPICE, 2013–2015) developed a uniform post-processing method for defining and quantifying precipitation events (Nitu et al., 2018). The process includes calculating a 30 min bucket-weight differential using thresholds and filters, effectively producing what was termed the Site Event Dataset (SEDS). For an event to be identified, the net precipitation duration needed to be sufficiently long (as measured by a precipitation detector or disdrometer), and the total accumulation (as measured by the reference automated weighing gauge) needed to be equal to or greater than a defined threshold (set at 0.25 mm when a reliable precipitation detector was available). This process was effective at creating a high-confidence data set for developing and testing transfer functions (Kochendorfer et al., 2017b) but, because of the rigorous filtering of shorter and smaller events, was not an effective means of filtering a time series.

The US Climate Reference Network (USCRN) uses the redundancy of the Geonor T-200B three vibrating-wire load sensors in the determination of precipitation events (Leeper et al., 2015). Initially, a pairwise calculation was used which relies on pairwise agreement of bucket-weight changes using the wire redundancy as a check on the measurement. This was determined to be sensitive to gauge evaporation and noise, leading to the development of a weighted average calculation using the change in bucket weight between successive sub-hourly periods for each transducer output. A weighted mean is then used to average the bucket weights, with greater weight given to less noisy measurements.

The Meteorological Service of Canada currently implements a real-time threshold filter in their data loggers to automatically determine the occurrence of precipitation events. The filter is based on the 15 min differential in the Geonor T-200B bucket weight (Mekis et al., 2018). Although this filter is unnamed, we call it the operational 15 min filter (O15) automated processing technique. This technique is included in this analysis and is described below in more detail. The filter tends to fail when the noise threshold is exceeded, resulting in false precipitation reports, and when evaporation exceeds the acceptable limits.

Limitations in the O15 technique led to the development of the neutral aggregating filter (NAF), previously known as “Brute Force” (Pan et al., 2016). The NAF, described in greater detail by Smith et al. (2019), iteratively adds all negative and small positive changes to proximate positive changes until all changes exceed a user-specified threshold. Because the technique preserves the total change in bucket weight over the time series, it cannot account for the negative drift that results from evaporation. To overcome this deficiency, the supervised neutral aggregating filter (NAF-S) was created to allow user intervention and minimize evaporation errors through interactive manual adjustment. Both NAF and NAF-S are explained in greater detail in the next section.

To overcome the limitations of the O15, NAF, and NAF-S techniques, we evaluated a moving-window modification of the NAF, implementing the NAF on 24 h overlapping windows, which we will call the segmented neutral aggregating filter (NAF-SEG). The objective was to obtain a robust post-processing technique that is completely automated; easily implemented; and successfully eliminates varying levels of noise, diurnal oscillations, and evaporation without significantly impacting the timing and amount of precipitation. This study introduces the NAF-SEG technique and examines its performance compared to the O15, NAF, and NAF-S methods.

Two data sources, both with 1 min resolution, were used to evaluate the O15, NAF, NAF-S, and NAF-SEG precipitation filters: a control (pre-processed) precipitation time series which is free of noise and drift and raw (minimally filtered) automated weighing-gauge data collected at a number of international sites, which contain varying levels of noise, diurnal oscillations, and evaporative drift. The control, pre-processed time series were used to evaluate all four filters – by adding synthetic noise, diurnal oscillations, and evaporative drift and then evaluating the ability of the filters to recover the original time series. The raw time series, following quality control procedures, were passed through each of the filters, and the supervised NAF-S output was used as the standard against which to evaluate the others.

Both data sources, raw data with real-world noise and control data with synthetic noise added, have advantages and disadvantages in assessing filter performance (Peters et al., 2014). Clean data with added noise provide a known “true” control but add the risk that the added noise and drift may not adequately capture the characteristics of real-world measurements. Raw measurements preserve observed noise patterns and capture the variability in noise behaviour across sites and instruments but do not provide a control time series for filter evaluation. By using both complementary data sources, we exploit their respective strengths and thus better assess the relative effectiveness of each filter.

This study evaluated four filters for processing the outputs of accumulating automated weighing precipitation gauges: three that were fully automated (O15, NAF, and NAF-SEG) and one that required manual supervision (NAF-S). Overall, NAF-S and NAF-SEG outperformed O15 and NAF; both NAF-S and NAF-SEG showed similar skill in compensating for evaporative losses and eliminating false detections caused by random noise and diurnal oscillations. O15 performed well in low-noise cases with minimal evaporation but generated false precipitation detections when the data were noisy and often underestimated evaporative losses. NAF performed well in cases with minimal evaporation regardless of the noise level but did not correct for evaporative losses. NAF-SEG performed consistently well and provided a fully automated alternative that matched the skill of the manual NAF-S method. Moreover, NAF-SEG added a direct estimate of evaporation, without the user intervention required by NAF-S or the 1 mm threshold required by O15. Similar evaporation estimates are not directly available from the other techniques.

Although NAF-SEG did not perfectly recover the synthetic evaporation that was added to the control time series (the recovery rates were 81 % to 116 % depending on the noise level), it performed as well as the manually supervised NAF-S technique. Both NAF-S and NAF-SEG failed to disentangle precipitation and evaporation when they occurred on the same day. The challenge to do so may be insurmountable. The imperfect recovery of synthetic evaporation, coupled with the sensitivity of the recovered evaporation to noise, highlights the need to implement measurement protocols that minimize evaporative losses. We recommend the use of NAF-SEG as a screening technique to identify gauges and locations that have significant evaporative losses and then to implement adequate measures to minimize those losses, such as modifications to the oil and antifreeze mixture used to prevent freezing and evaporation.

Overestimation of precipitation by the O15 method occurs when the noise exceeds the filter's prescribed threshold of 0.2 mm. This value for the threshold has been set based on experience as a necessary and calculated balance between eliminating real precipitation events and detecting false events. When the noise level is low, as in the low-noise scenario of the control data, the O15 technique works successfully. However, noise patterns vary substantially from site to site and among gauges, as illustrated by Nitu et al. (2018), and often exceed the filtering capabilities of O15. It should also be noted that the unprocessed data in our tests were pre-filtered using a Gaussian filter with a 4 min window, which was integrated into the SPICE quality control process prior to testing the algorithms. This likely resulted in the O15 performing better than it would have in the operational setting, but this was not confirmed.

The NAF technique is fundamentally effective at filtering noise and diurnal oscillations but underestimates precipitation when evaporative losses occur because the algorithm forces the precipitation total to match the final raw bucket weight in the time series, with evaporation assumed to be zero. The NAF-SEG technique, which implements NAF over 24 h windows, maintains all the strengths of NAF with the added functionality of automating the detection and removal of bucket evaporation. Neither NAF-S nor NAF-SEG removes evaporation perfectly, particularly when it occurs in combination with precipitation, but both represent a major step forward compared to other processing methods. We attribute the effectiveness of NAF-SEG to two characteristics of precipitation events: first that evaporation is relatively small during periods with precipitation and second that both precipitation and evaporation are persistent over timescales of days. In the development of NAF-SEG, a 24 h moving window was chosen to minimize the impact of temperature-related diurnal oscillations, but fortuitously the 24 h window also served to separate days with precipitation and little evaporation from days with evaporation and little or no precipitation. The performance of NAF-SEG may decline when signal noise is due to non-cyclical temperature fluctuations, such as those that occur during strong synoptic events. Although this possibility was not assessed, it is one that users should be aware of.

As mentioned in the introduction to NAF-SEG, a sensitivity analysis was performed for a range of P* values from 0.0001 to 0.5 mm using the pre-processed high-noise time series for both warm and cold seasons. The analysis showed negligible sensitivity as P* ranged from 0.0001 to 0.05 and higher sensitivity as P* further increased to 0.5 mm for both seasons. Given the relative insensitivity of NAF-SEG to P* < 0.5 mm, the use of 0.001 mm seems to be an appropriate baseline value for both seasons; users may want to further experiment with the parameter as their own data require.

NAF-SEG provides an attractive alternative to NAF when negative evaporative drift is present in the raw data, but it is not designed to handle all contingencies. For instance, unexplained positive then negative excursions in bucket weight are sometimes observed. If the positive and negative excursions are separated by more than 24 h (the size of the window), the NAF-SEG will errantly attribute the positive excursion to precipitation and the negative excursion to evaporation.

The results of the testing on unprocessed time series from different sites, seasons, and gauge configurations showed that NAF-SEG generally outperformed O15 in both cold- and warm-season test cases. Of the 44 cold-season test cases, O15 outperformed NAF-SEG in only two cases: the DFAR and unshielded Pluvio2 gauges at WFJ (2016–2017). However, these gauges may not have been serviced adequately; note the extreme evaporation rates as evidenced in the high biases between NAF and NAF-S in Table A1. This diminishes their usefulness for this evaluation; they were among the most challenging to process, with the greatest uncertainty in the supervised NAF-S output that served as the reference standard.

Filter evaluation was more limited in the warm season because the raw site data were obtained from the SPICE project, which focused on the measurement of solid precipitation. Still, we were able to assemble 11 warm-season cases. The warm-season data were expected to differ from the cold-season data in two respects: higher evaporative losses and different noise characteristics. Each of the filters generated a higher RMSD in the warm season than the cold season; the greatest increase was found for O15, consistent with the pre-processed control experiments. In general, NAF-SEG outperformed both NAF and O15 in the warm season. NAF-SEG outperformed O15 in all warm season cases for r and RMSD and resulted in a lower or similar seasonal bias in 7 of the 11 cases. The NAF-SEG totals consistently underestimated warm-season precipitation, but the biases were small, averaging 1.7 % compared with 1.0 % for the cold season. Regardless of the sample size, the performance metrics all show that NAF-SEG outperformed both NAF and O15 in the warm season as well as the cold season.

The evaluation of filter performance based on raw site data begs the following question: how reliable are the NAF-S outputs as reference standards, given that they rely on the operator's subjective judgement during the interactive elimination of negative drift and other spurious bucket-weight changes? We acknowledge that operator bias is possible but are confident that its impact in this study is minimal. A single, skilled operator processed all of the data and made every attempt to apply the NAF-S method consistently. Adding further confidence to the NAF-S outputs are the tests with control data, which independently demonstrated the efficacy of the NAF-S to eliminate noise and evaporative drift.

One suggestion to improve the quality of data from accumulating precipitation gauges is to add disdrometers, which detect the current weather conditions, to the site measurements and then incorporate their outputs into the quality control and filtering process. These augmented observations could be used to refine the noise filtering by automating the high-temporal-resolution (e.g. 1 min) detection of light precipitation events and assist in removing false precipitation detections. These ancillary data were used in this way during SPICE (Nitu et al., 2018) and should be further explored for enhancing operational filtering.

This study reports the development and implementation of a robust, fully automated technique for post-processing data from automated weighing precipitation gauges. The NAF-SEG technique is designed to eliminate varying levels of random noise and diurnal oscillations as well as correcting for negative drift from bucket evaporation. An intercomparison of four filtering techniques shows that the O15, although simple and deployable in real time, fails when noise levels exceed the filter's threshold and may undercompensate for bucket evaporation. NAF, although highly effective in eliminating noise, does not correct for evaporative losses. NAF-S, which adds manual supervision to NAF, is effective in removing noise, eliminating spurious data, and correcting for negative drift from evaporation. However, it is labour-intensive and best suited to complete seasonal time series.

Our results show that NAF-SEG is equally as effective as NAF-S in eliminating noise and evaporative drift from automated weighing-gauge precipitation measurements. When tested against a control data set with added synthetic noise and evaporation, NAF-SEG was able to recover the original control to within ±3 % of the total, with a lower RMSE than the other techniques. When evaluated on 55 raw time series from various sites, years, and gauge configurations, NAF-SEG outperformed O15 and NAF and gave the highest mean correlation coefficient and lowest mean RMSD.

One limitation of NAF-SEG is that it requires 24 h data segments; consequently, it cannot be deployed for real-time processing of automated weighing-gauge precipitation measurements. Until other alternatives are found, we recommend the use of a simple threshold filter like O15 for real-time applications, but with the archiving of the raw 1 min time series for subsequent enhanced quality control, reprocessing using NAF-SEG, and the archiving of the NAF-SEG outputs. This, in combination with routine site servicing to minimize evaporation and other sources of noise, can result in improved operational precipitation data.