The U.S. Cooperative Observer Program (COOP) network was formed in the early 1890s to provide daily observations of temperature and precipitation. However, manual observations from naturally aspirated temperature sensors and unshielded precipitation gauges often led to uncertainties in atmospheric measurements. Advancements in observational technology (ventilated temperature sensors, well-shielded precipitation gauges) and measurement techniques (automation and redundant sensors), which improve observation quality, were adopted by NOAA’s National Climatic Data Center (NCDC) into the establishment of the U.S. Climate Reference Network (USCRN). USCRN was designed to provide high-quality and continuous observations to monitor long-term temperature and precipitation trends, and to provide an independent reference to compare to other networks. The purpose of this study is to evaluate how diverse technological and operational choices between the USCRN and COOP programs impact temperature and precipitation observations. Naturally aspirated COOP sensors generally had warmer (+0.48°C) daily maximum and cooler (−0.36°C) minimum temperatures than USCRN, with considerable variability among stations. For precipitation, COOP reported slightly more precipitation overall (1.5%) with network differences varying seasonally. COOP gauges were sensitive to wind biases (no shielding), which are enhanced over winter when COOP observed (10.7%) less precipitation than USCRN. Conversely, wetting factor and gauge evaporation, which dominate in summer, were sources of bias for USCRN, leading to wetter COOP observations over warmer months. Inconsistencies in COOP observations (e.g., multiday observations, time shifts, recording errors) complicated network comparisons and led to unique bias profiles that evolved over time with changes in instrumentation and primary observer.
The detection and attribution of climate signals often rely upon manually operated networks with lengthy records (Rasmussen et al. 2012). In the United States, the Cooperative Observer Program (COOP) network has collected many decades of manual observations from thousands of stations. While COOP was originally designed to provide daily data to agricultural communities (National Research Council 1998; Daly et al. 2007; Fiebrich and Crawford 2009), it has since evolved into the backbone of the U.S. climatology dataset and is widely used in national (Melillo et al. 2014; Karl et al. 2009), regional (Allard et al. 2009), and local (Diem and Mote 2005) climate assessments. Observations of temperature and precipitation from the COOP network are taken once daily from naturally ventilated temperature shields and unshielded precipitation gauges, which can introduce systematic biases based on prevailing meteorological conditions. In addition, changes in COOP instrumentation and shielding, time-of-observation inconsistencies, and station moves affecting sensor exposure (Peterson et al. 1998; Pielke et al. 2007) have occurred throughout the network’s history. These and other factors can obscure climatic trends (Peterson and Vose 1997; Peterson et al. 1998), requiring homogenization corrections (Easterling et al. 1999; Menne and Williams 2009; Menne et al. 2009) that are currently available at monthly time scales.
The U.S. National Research Council (NRC 1998), citing these and other challenges (Karl et al. 1995; Goodison et al. 1998) and in alignment with more recent findings (Fiebrich and Crawford 2009; Hubbard et al. 2004; Allard et al. 2009; Daly et al. 2007), identified the need for a network that was specifically designed to monitor climate and support climate observation efforts in the United States. Drawing upon the successful technological innovations that have helped fuel widespread adoption of automated networks for consistency (Fiebrich 2009; Fiebrich and Crawford 2009) and redundant (three sensors) measurement techniques that improve data quality and continuity (Karl et al. 1995), NOAA’s National Climatic Data Center (NCDC) deployed an automated climate observation network named the U.S. Climate Reference Network (USCRN; Diamond et al. 2013).
Designed with the purpose of detecting U.S. climate trends, the USCRN program broke ground on its first station in 2000. USCRN has deployed to date 132 stations in stable, representative locations across the conterminous United States (114), Alaska (16), and Hawaii (2) (Diamond et al. 2013). Observations of temperature and precipitation are taken from precision instruments calibrated to National Institute of Standards and Technology (NIST) traceable standards and regularly monitored and maintained. To preserve data quality and continuity, temperature and precipitation sensors are well shielded and installed redundantly as previously noted. Station observations are transmitted hourly via satellite to NCDC, where data are processed through quality control (QC) systems and made available over the web with a latency of one to several hours (Diamond et al. 2013).
As the longevity of the USCRN observation record increases, network design differences (instrumentation, shielding, and observation methods) between COOP and USCRN will become increasingly relevant in future climate assessments. Network design has been found to impact observations of both temperature (Guttman and Baker 1996; Lin and Hubbard 2004) and precipitation (Rasmussen et al. 2012). Lin and Hubbard (2004) identified sensor sensitivity, analog signal conditioning, and data acquisition (e.g., datalogger) as sources of differences in temperature measurements between USCRN, COOP, and other networks. In addition to potential biases associated with station exposure (Guttman and Baker 1996), Harrison (2011) and Nakamura and Mahrt (2005) also noted that during calm conditions naturally aspirated thermistors (as used in the COOP network) observe warmer daily maximum temperatures due to radiative loading compared to well-ventilated (aspirated) shields (installed at USCRN stations). A similar radiative bias occurs during the evening hours when radiative cooling results in cooler nocturnal temperatures from naturally ventilated sensors. The albedo of the underlying ground cover can also impart additional radiative imbalances during daylight hours when solar rays reflected from the ground can enter temperature shielding and impact observations (Hubbard et al. 2001). Reflective radiation biases can change seasonally with snow cover and health of the underlying vegetative ground cover (green-up and brown-down). For precipitation, Rasmussen et al. (2012) and others (Devine and Mekis 2008; Goodison et al. 1998; Groisman and Legates 1994; Nitu and Wong 2010; Sevruk et al. 2009) identified wind conditions, gauge wetting factor (buildup of precipitation on gauge orifice surfaces), gauge evaporation (loss of precipitation from gauge prior to measurement), and orifice blockage as some of the main factors contributing to uncertainties in precipitation measurements (Sevruk et al. 2009). When surface wind deflects around a precipitation gauge, it can create flow patterns that inhibit precipitation from falling into the gauge (Sieck et al. 2007). The impact of wind on precipitation catch is more pronounced for less aerodynamic hydrometeors, such as snow, with errors up to 50% of total precipitation possible (WMO 2008; Sevruk et al. 2009). The use of shields around precipitation gauges, as in USCRN, disrupts such flow patterns over the gauge opening, allowing hydrometeors to fall into the gauge and reduce measurement errors due to winds. This impact is more discernable for frozen hydrometeors. Groisman et al. (1999) in their comparison study of COOP and automated Fisher and Porter gauges found that gauge exposure to wind did not “noticeably affect” observation error for liquid hydrometeors.
Human observers also contribute to observational uncertainty. Inconsistencies in time of observation, multiday reports, and observation methods are other sources of error from manually operated networks. Conversely, automated networks if provided adequate power and excellent maintenance can preserve consistency by ensuring measurements are taken at designated times and in the same way (Fiebrich and Crawford 2009). Over time the quality of observations taken from both automated and manual networks can degrade with sensor health (age of sensor), dirt buildup, and pest infestations, which are more difficult to detect and resolve for automated networks given the lack of daily visits. However, USCRN redundant sensors will help mitigate some of these issues.
Earlier investigations (Hubbard et al. 2004; Sun et al. 2005) comparing USCRN with other networks [COOP, Automated Surface Observing System (ASOS)] primarily focused on side-by-side measurements of temperature often excluding precipitation disparities. In addition, the side-by-side setup such as studied in Hubbard et al. (2004) concentrated on instrumentation biases related to sensor shielding, which can change with prevailing meteorological conditions. However, there are additional network related biases that can arise from maintenance (e.g., degradation of aging sensors, calibration standards), measurement practices (e.g., reporting times, changing observers), and local site characteristics not captured in those studies. The combined effects of sensor-, maintenance-, measurement-, and siting-related biases may be more useful to studies using daily COOP measurements in addition to efforts to homogenize the daily record. The purpose of this study is to compare observations of temperature and precipitation from nearby members of USCRN and COOP networks under normal operating conditions.
The COOP network currently consists of approximately 8000 stations across the United States (National Weather Service 2014b). Volunteers take maximum and minimum temperature and precipitation observations once daily at designated times, which vary from station to station. Maximum and minimum temperatures have traditionally been observed from liquid-in-glass (LiG) thermometers shielded in naturally aspirated cotton-region shelters (CRS) (Fig. 1a). However, these thermometers have gradually been phased out since 1983 in favor of digital thermistor systems known as the maximum–minimum temperature sensor (MMTS), which now make up more than 90% of COOP stations where metadata are available. The MMTS thermometer has an operational range between −62° and 60°C (−80° and 140°F) and is equipped with an indoor digital display. The original display had to be reset daily; however, Nimbus displays have the capacity to store up to 35 days of previous observations, including hourly measurements (Sensor Instruments Co. Inc. 2000). In addition, these temperature sensors are shielded from solar insolation using a multiplate shield (Fig. 1a).
Precipitation observations are taken from unshielded 8-in. gauges that require a calibrated dipstick to measure precipitation in an internal concentrator tube (Fig. 1b). Daily accumulations can be measured to a precision of one-hundredth of an inch or 0.3 mm. Frozen precipitation measurements are reported as ground accumulation and liquid equivalent, which require additional steps. If snow is anticipated, the central tube and funnel top are removed, so as to allow snow to fall naturally into the outer diameter without interference. The snow is then melted and poured into the concentrator tube and measured. After each observation, observers must empty the gauge for the next 24-h period. More information regarding COOP observations and standards can be found in the National Weather Service 10-1315 (National Weather Service 2014a) manual.
USCRN stations are automated with observations of temperature and precipitation taken at a subhourly frequency. Temperature is observed using platinum resistance thermistors (PRTs), each of which is housed within separate fan-ventilated Met One shields (Fig. 1c) with an operational range of −60° to 85°C (−76° to 185°F). While observations from redundant temperature sensors are available, USCRN official temperature, which is compared with COOP in this study, is a single value derived from the redundant observations that passes a series of QC checks.
Precipitation is observed with a Geonor T-200B weighing-bucket-type gauge located within a small double fence intercomparison reference (SDFIR) shield surrounding a single Alter shield at most stations (Fig. 1d). The all-weather gauge utilizes vibrating wire technology to detect precipitation events as small as 0.2 mm with a precision of 0.1 mm (Baker et al. 2005; Duchon 2008). During winter, an antifreeze mixture is added to the gauge reservoir to melt frozen hydrometeors (only the liquid equivalent is reported) and an orifice heater is activated to prevent snow and ice from accumulating along the rim of the gauge. Similar to temperature, the redundant measures of gauge depth are processed through a series of QC checks before they are used to report official precipitation. Because it is possible for noise (small up and down depth variations) to occur coincidentally among the redundant sensors, an auxiliary disdrometer (the Vaisala DRD11A) is used to distinguish between dry and precipitating subhourly periods. USCRN stations are also equipped to monitor additional variables, including solar radiation, surface wind speed, surface IR temperatures, relative humidity, and soil moisture and temperature at 5-, 10-, 20-, 50-, and 100-cm depths. In this study, network differences will be discriminated by surface wind speed, which is observed at a height of 1.5 m using the Met One model 014A three-cup anemometer with mean wind speed and 10-s gusts provided hourly. For additional information on USCRN and its monitoring of surface and subsurface conditions, refer to Diamond et al. (2013) and Bell et al. (2013).
To ensure observational differences are the result of network discrepancies, comparisons were only evaluated for stations pairs located within 500 m. This distance criterion was selected based on an analysis by Guttman and Baker (1996). Thirteen station pairs met this criterion with collocated station pairs ranging in distance from 42 to 400 m. However, preliminary examination of station metadata showed the St. Mary, Montana, pair was sited within close proximity of buildings, overhanging trees, and other obstacles, which would have skewed daily comparisons (temperature and precipitation) given the small sample size and therefore excluded from the analysis. The remaining 12 station pairs (see Fig. 2) were dispersed across the contiguous United States for both northern and southern latitudes, providing network comparisons from mountainous (Arco, Idaho, and Dinosaur, Colorado) to coastal (Kingston, Rhode Island) and desert (Las Cruces, New Mexico) to humid (Holly Springs, Mississippi) locations.
To compare the two networks, USCRN subhourly data were aggregated into 24-h periods to match daily COOP measurements at the designated observation times. USCRN daily maximum and minimum temperatures were set to the warmest and coolest 5-min average taken every 10 s throughout the 24-h period. Prior to 2006, the moving average was not available and therefore the maximum and minimum are simply the warmest and coolest 5-min clock averages (1–5, 6–10, …, 56–60 min of the hour), respectively. Daily precipitation is reported as a 24-h accumulation. Start and end hours for USCRN daily aggregates were defined by COOP observation time as noted in Table 1 with additional adjustments accounting for daylight saving time when appropriate. For three of the stations [Holly Springs (4N), Agate (3N), Nebraska; and Stillwater (2W), Oklahoma], COOP observation times changed during the study period with USCRN aggregation times appropriately adjusted. In addition, daily aggregates were only computed if no more than one hour of data was flagged or missing; otherwise, not available (N/A) would be recorded. On days with missing data, USCRN and COOP observations were both assigned N/A if either network had missing daily data. These efforts were taken to ensure a more straightforward comparison.
USCRN daily maximum and minimum temperatures on average differed from neighboring COOP stations by −0.48° and +0.36°C, respectively. The more moderate USCRN daily extremes resulted in a smaller diurnal temperature range (DTR) that was on average 0.84°C less than COOP stations. The magnitude of temperature differences were on average larger for stations operating LiG systems, with mean maximum (minimum) temperature differences of −0.66°C (+0.48°C), compared to −0.39°C (+0.31°C) differences for the MMTS system. Part of the reduction in network biases with the MMTS system is likely due to the smaller-sized shielding that requires less surface wind speed to be adequately ventilated.
While overall mean differences were in line with side-by-side comparisons of ventilated and nonventilated sensors (Hubbard et al. 2004), there was considerable variability in the mean magnitude of network differences for maximum and minimum temperatures from station to station (Fig. 3). The size of network differences varied regardless of sensing technology (LiG or digital) and station separation distance. However, there was some consistency in the direction of network differences at least for maximum. For all station pairs, COOP stations observed warmer maximum temperatures on average. The direction of network differences was less consistent for minimum temperatures, where COOP stations had warmer (4 of 12) and cooler (8 of 12) daily minima compared to USCRN on average. The variability in magnitude (for maximum and minimum) and direction (for minimum) of temperature differences among station pairs may be explained by differing meteorological conditions (surface wind speed, cloudiness), local siting (heat sources and sinks), and sensor (poor calibration, health)/human (varying observation time, reporting error) error as noted in Wu et al. (2005).
To further explore the impacts of prevailing meteorological conditions on network comparisons, temperature differences were categorized by wind speed (Figs. 4a–d). While there were daily outliers greater than ±20°C for both maximum and minimum temperatures, the range in network differences for maximum and minimum temperatures visually seemed to reduce with increasing wind speeds (Figs. 4a and 4c). For maximum temperature, mean network differences reduced with increasing wind speed as sensor shielding became better ventilated (Fig. 4b). However, minimum temperature differences increased slightly with wind speed (Fig. 4d), which according to Fig. 4c may be related to an imbalance of more large positive than large negative differences on windy days.
Under calm conditions one might expect radiative imbalances between naturally and mechanically aspirated shields or differing COOP sensing technologies (LiG vs digital sensors) to drive network differences (Gall et al. 1992; Quayle et al. 1991). While these biases may account for some variability in the magnitude of network differences, they do not explain the variation in the sign of minimum temperature differences (Fig. 3). Siting characteristics, which can change over short distances, have been found to influence surface temperatures (Guttman and Baker 1996). For instance, stations located near trees (Crossville, Tennessee, and Gaylord, Michigan) compared to those in the open can be subject to an insulating effect as described by Groot and Carlson (1996) and result in warmer daily minima. Likewise, small elevation changes between USCRN and COOP stations (Dinosaur) or those located near a gravel parking lot (Murphy, Idaho) can result in sharp thermal contrasts over short distances. The impacts of siting are likely more pronounced during calm conditions when localized factors have more time to affect the thermal properties of the overlying air mass. This is particularly true during evenings when the boundary layer decouples and collocated station pairs are no longer sampling a homogeneous air mass (Yao and Zhong 2009).
At higher wind speeds (≥2 m s−1), any potential effects of local land cover/terrain appeared to diminish (smaller envelope of variation) with improved mixing and radiative imbalances becoming more dominate as COOP observations become more consistently cooler than USCRN in-line with side-by-side comparisons. However, large outliers (>20°C) remained, which may have biased mean differences in a positive direction for the stronger wind categories. These results suggest that some of the minimum temperature differences during windier conditions may not be related to instrumentation (radiative) biases, but perhaps linked to variations in observation times or other observer errors. In these cases, wind speed at the time of USCRN minimum temperature may not correspond well with the wind speed at time of COOP minima. In addition, multiday or time-shifted observations associated with periods of rapid temperature change (i.e., frontal passages) can result in large network differences.
Over monthly time scales, there was little seasonal variation in network biases for both maximum and minimum temperatures on average; however, some station pairs did have monthly trends that were more pronounced for maximum than minimum temperatures. Easterly stations (i.e., Holly Springs), for instance, had larger maximum and minimum biases on average over summer than winter months (Figs. 5a and 5b). The greater network biases over summer months were thought to be mostly attributed to increases in solar radiation (radiative biases) in addition to changes in seasonal wind patterns, resulting in fewer well-aspirated windy days. Some stations located over the complex terrain of the northwest (i.e., Dinosaur) had an opposite trend with the magnitude of network biases lessened over warmer months; however, this was not true with all northwestern station pairs. Still other stations (those in the desert Southwest such as Muleshoe, Texas) had little or no change from month to month despite well-known seasonal winds from the North American monsoon. The exact causes of such seasonal variations or lack thereof are not completely understood given the limited number of stations considered in this study. However, some causes to seasonal variations can include seasonal patterns in surface winds that impact shielding ventilation, changes in ground cover albedo (snow cover, vegetation green-up or brown-down) that influence reflection of solar radiation from the ground into senor shielding, or changes in solar zenith angles for higher latitude stations.
Over annual time scales, mean maximum and minimum temperature biases were also fairly constant over time (Figs. 6a and 6b) with the exception of a few stations. For instance, the Arco station pair had one of the largest annual changes in temperature biases between 2007 and 2008 for both maximum and minimum temperatures. Station metadata over this period reveal a change in Nimbus serial numbers (instrumentation replacement) that coincided well with sizeable shifts in daily temperature differences, possibly due to calibration difference between Nimbus units (Figs. 7a and 7c). Other station pairs had apparent shifts in temperature bias that were correlated with other types of COOP metadata events. In late 2009, the frequency of large temperature differences at Dinosaur was considerably reduced at about the time a new primary station observer was appointed. When this primary observer departed in late 2011, large temperature differences reemerged (Figs. 7b and 7d). Variations over annual time scales identified the sensitivity of network differences to changes in station observer and impacts of sensor error, which may be difficult to identify without collocated stations or redundant sensors.
Daily precipitation observations were not well correlated between the two networks. Scatterplots of daily observations (Fig. 8a) indicate the two networks were not temporally aligned given the clustering of points along the x and y axes. In addition, rare recording errors such as duplicate observations, decimal misplacement, misplacement of monthly total for last day of the month, and multiday sums (when COOP observations are based on multiple days of precipitation but entered on one calendar day) resulted in large network contrasts. These inconsistencies and other errors made it challenging to accurately quantify network differences at the daily scale.
To proceed, daily observations from both networks were grouped into precipitation events (multiday accumulation) to improve the temporal alignment. Each event begins on a day when either network observes precipitation and ends on the first day neither reports precipitation. Precipitation events were discarded if either network reported missing data during an event. By aggregating daily data into multiday events, most timing inconsistencies (shifts of 24 h or less) were resolved. From an event perspective, precipitation observations from the USCRN and COOP station pairs were better correlated as shown in Fig. 8b; however, shifts in observation time greater than 24 h (multiday observations) can still complicate comparison results.
Overall, USCRN observed 1.5% less precipitation than neighboring COOP stations. However, this was not uniform across all station pairs by year or by season (Figs. 9a and 9b). Annual differences shifted from dryer COOP observations (2005 and 2006) relative to USCRN to slightly wetter from 2007 onward. Seasonally, COOP reported less (more) precipitation than USCRN stations during winter (summer) months. The dryer wintertime COOP observations are likely due to the lack of gauge shielding but may also be impacted by the added complexity of observations (winterizing and melting). As noted by WMO (2008), wind-induced errors are more pronounced for frozen (10%–50%) than liquid (2%–10%) hydrometeors. These results may also help to explain some of the annual variations, as more of the southern stations were included in the analysis after 2007. One exception to this was the Gaylord pair, where the COOP station reported more winter precipitation. An analysis of depth change revealed USCRN algorithms used to calculate depth change (precipitation) underreported precipitation at this station due to a suboptimally performing disdrometer, which at times improperly classified precipitation as sensor noise.
COOP reported precipitation amounts over warmer months were greater than USCRN on average. These differences may be partially related to the enhanced spatial variability of unorganized convective activity that is more dominant over warmer months (Tokay et al. 2014). This is particularly true for coastal areas (Kingston) or in the Southeast, where daytime convection could trigger afternoon pop-up thundershowers. In one particular event at Gaylord, the COOP observer reported 20.1 mm more than the USCRN gauge located within 133 m. Despite radar estimates that were more similar to USCRN measurements, there were nearby areas with intense precipitation, suggesting the COOP observations may have been valid.
Wetter COOP observations over warmer months may also have been associated with seasonal changes in gauge biases. For instance, observation errors related to gauge evaporation and wetting factor are more pronounced in warmer than cold conditions, while the opposite is true of wind-related errors (WMO 2008), as noted previously. Because of design, the Geonor gauge is more prone to wetting errors with a larger wetting factor (9.0) than the standard 8-in. gauge used at COOP stations, which had a wetting factor of 3.2 (WMO 2008; M. Hall 2014, personal communication). In addition, USCRN does not use an evaporative suppressant to limit gauge evaporation during the summer, which is also not an issue for the funnel-capped COOP gauge according to Golubev et al. (1992). The combination of elevated biases for USCRN and reduction in COOP wind-related errors over warmer months might help to explain a portion of the seasonal variations in network differences.
To investigate these results further, precipitation events were categorized by air temperature, wind speed, and precipitation intensity (Fig. 10). Only events where both networks observed precipitation (neither station reporting zero for an event) were considered, similar to Tokay et al. (2010). USCRN hourly temperatures during periods with precipitation were averaged into an event mean and used to group events into warm (mean temperature ≥ 5°C), near-freezing (mean temperature between 0° and 5°C), and freezing (mean temperature < 0°C) conditions. For all warm and near-freezing events, USCRN observed 4.4% and 2.9% less precipitation than nearby COOP gauges, respectively (Fig. 10a). This tendency was reversed for frozen conditions, when USCRN observed 10.7% more precipitation than COOP. Shielded USCRN gauges observed more precipitation for frozen hydrometeors in part due to shielding, as suggested by Rasmussen et al. (2012).
Precipitation event differences were also impacted by wind speed and intensity. Wind speed criteria were based on Guttman and Baker (1996) and reported in meters per second as an average over hours USCRN observed precipitation. For light wind events (≤1.5 m s−1), USCRN reported 7.5% less precipitation than COOP. As wind speeds increased, USCRN accumulation deficit was reduced to 2.8% less than COOP for moderate winds with USCRN exceeding COOP by 3.4% for wind speeds of 4.6 m s−1 or greater (Fig. 10b). As with temperature, there was a change in sign of network differences when the impact of gauge shielding was more influential on precipitation capture.
For precipitation intensity, USCRN observed less than neighboring COOP for all categories (Fig. 10c). However, COOP precipitation observations were more similar to USCRN during higher rate events (≥4 mm h−1), with COOP having a 0.33% wet bias. This was considerably less than the 3.8% wet bias for the lightest intensity (≤2 mm h−1) category. The reduction in precipitation differences with intensity can be partially explained by a reduction wind and wetting loss biases with raindrop size. Intense precipitation events consist of larger hydrometeors that are less aerodynamic and reduce wind biases from COOP gauges (Sieck et al. 2007) and quickly moisten gauge orifice walls, cutting down on wetting losses from the Geonor gauge.
On average, temperature differences between USCRN and COOP were in general agreement with side-by-side comparisons conducted by Hubbard et al. (2004) of aspirated and nonaspirated thermometer systems, with naturally ventilated systems observing warmer daily maximum and cooler daily minimum temperatures. In addition, mean temperature biases were reduced with the smaller-sized shield of the MMTS sensor as noted in Quayle et al. (1991).
However, comparisons from an operational setting revealed that not all station pairs had similarly sized mean temperature differences. The variability in the magnitude of temperature bias was detected for COOP stations operating both LiG and digital sensors, suggesting other factors might have influenced mean differences. For instance, prevailing meteorological conditions can at times amplify (clear skies) and reduce (windy conditions) the magnitude of radiation errors of naturally aspirated systems used in the COOP network. These meteorological conditions can vary from day to day and seasonally (i.e., frequency of synoptic activity) in addition to geographical variations (high latitudes, maritime, mountainous) that were not fully explored in this study.
Additionally, ground cover and local siting may be other factors contributing to the variation in magnitude and direction (for minimum) of network biases among station pairs. Hubbard et al. (2001) convincingly showed that the reflective properties of the underlying ground cover can contribute to radiative errors during the day by reflecting incoming solar radiation into the sensor housing of both CRS and mulitplate shields used in the COOP network. The reflective properties of the underlying ground for northern snowy stations may change over submonthly and seasonal scales depending on snowpack. Other stations in the network may also have changes in surface albedo throughout the year with vegetative green-up and brown-down.
The effects of local siting seemed to be more pronounced when the atmosphere was less mixed, affecting primarily minimum temperatures of stations located in proximity of trees (Crossville and Gaylord) and gravel pads in the case of Murphy. For maximum temperature, the impact of local siting was not as discernable, which may be due to daytime boundary layers generally being well mixed at time of maxima. While the effects of ground cover and local siting were not fully explored here, these factors when combined with prevailing meteorological conditions may help to partially explain the variations in magnitude (maximum and minimum) and direction (minimum) of mean network temperature differences among station pairs in addition to seasonal variations.
Annual evaluations of temperature differences revealed inconsistencies in observation practice (varying observation time, accounting errors, and others) that seemed to change with station observer at Dinosaur in addition to sensor-related differences at Arco. While accounting errors (decimal misplacement) can result in large outliers for temperature and precipitation, varying observation time and multiday observations are more difficult to detect because the network differences are less systematic (variable in magnitude and sign). For instance, the magnitude of network temperature biases caused by inconsistent reporting time will be amplified (lessened) over periods with a rapidly changing (persistent) air mass, which may also have a seasonal cycle depending on location. In cases of frontal activity, extreme differences (>20°C) in network temperature were observed. An example is shown in Fig. 11, where the observer seemingly reported a multiday minimum that was 20.9°C cooler than USCRN. The COOP minimum temperature was likely observed after the defined COOP observation time, following the passage of a cold front.
In this study, these scenarios complicated the analysis of temperature and precipitation event differences. This is particularly true when network differences were categorized by prevailing meteorological condition. If observations of daily extremes are not taken from the stated 24-h period, then the USCRN-monitored conditions may not reflect the actual atmospheric conditions when COOP measurements were taken, which may skew the results. In rare cases, large differences due to observer error (temperature and precipitation) were submitted to NCDC dataset stewards for corrective action and removed from this study. One possible solution to observation time inconsistencies might be to allow observers to report observation times as was noted in Tokay et al. (2010), rather than presuming all observations are taken at a designated time. Such metadata would greatly enhance the usability of the COOP network and likely would have improved comparisons between USCRN and COOP and better explain network biases with respect to atmospheric conditions, which would be more useful to daily homogenization efforts.
Despite important network differences for temperature and their possible causes as noted previously, it should be noted that differences between the two networks using very different sensors with different types of aspiration (natural vs mechanical) are expected. While such differences may affect meteorological evaluations (daily highs and lows), climate trends, presuming the types of errors remain consistent over time, may not be negatively impacted. For instance, comparisons between USCRN and homogenized COOP data (referred to as the U.S. Historical Climatology Network), which attempt to remove such inconsistences as described by Menne and Williams (2009), had very similar maximum and minimum national temperatures (Fig. 12), suggesting these routines are effective in line with Menne et al. (2010). However, these homogenization methods are currently not available at the daily scale, and this study identifies some of the day-to-day variations in network biases that will make daily homogenization efforts challenging without another network that can provide consistency at daily time scales, such as USCRN.
Precipitation differences between the two networks were primarily influenced by hydrometeor type and wind conditions. For frozen hydrometeors or strong wind conditions, USCRN well-shielded gauges reported more precipitation than unshielded COOP gauges. These results are in-line with other studies comparing shielded and unshielded gauges (Golubev et al. 1992). However, the opposite was true for liquid hydrometeors or in calm conditions when COOP stations tended to observe more precipitation than collocated USCRN gauges. While these results are in contrast to Golubev et al. (1992), they are similar to recent comparison studies using the Geonor T-200B gauge (Devine and Mekis 2008; Gordon 2003). Devine and Mekis (2008) reported that the Geonor gauge underreported a reference pit gauge more so than a manually operated type B gauge used in Canada. In addition, the magnitude of undercatch with the Geonor T-200B gauge with respect to the pit gauge varied with total precipitation (an approximate to precipitation intensity) with larger biases for smaller rain events similar to this study. However, it should be pointed that the Canadian study did not use the same algorithm to compute precipitation from the Geonor, as USCRN technicians have developed their own open source approach to processing Geonor gauge depths.
Differences between USCRN and COOP precipitation totals over warmer periods are likely linked to Geonor gauge design, maintenance practice, and computational methods used to evaluate precipitation, which can impact measurement errors. The Geonor T-200B gauge has an open vertical shaft instead of a funnel as used in the COOP network. The open shaft design, while minimizing splash-in and -out errors, increases the surface area fallen precipitation can adhere to, resulting in a wetting factor (9.0) that is nearly 3 times larger than the COOP gauge (3.2). However, when the funnel is removed for winterization, the COOP gauge will presumably have a similar wetting factor to the Geonor. In addition, the absence of a funnel for the Geonor allows evaporation to take place over warmer months, which according to Golubev et al. (1992) is negligible for the standard 8-in. gauge used at COOP stations. This is particularly true because no evaporative suppressants are added to the Geonor gauge over the warm season. It was thought that frequent observations at a subhourly temporal resolution would mitigate evaporation biases, as suggested by Sevruk et al. (2009). However, this does not imply that methods used to evaluate depth change (precipitation) would be completely insensitive to gauge evaporation. Additional analysis (not shown here) revealed evaporation may impact the start time and accuracy of USCRN event precipitation due to the use of a 2-h smoothing method. Given this potential sensitivity, USCRN computation methods used to determine depth change (precipitation) are currently being reevaluated with a goal of reducing evaporation and other algorithm sensitivities.
The combination of both wetting factor and evaporation create a seasonal bias cycle that is more pronounced over summer than winter months for the Geonor gauge. In a similar manner, the winterization of the COOP gauge (greater wetting factor) and the lack of shielding also results in a seasonal bias cycle that is amplified over winter months. The out-of-phase seasonal cycle of biases for the USCRN and COOP gauges may help to explain some of the seasonal differences between the two networks.
As noted previously with temperature, evaluations of network differences for precipitation by meteorological conditions can be complicated by observer inconsistencies. Despite grouping days with precipitation into multiday events, precipitation reports shifted more than 24 h can result in wetter COOP events. This mostly affects categorical differences because only events where both networks reported precipitation were used in the analysis shown in Fig. 10. An example of this is provided in Table 2, which shows USCRN and COOP daily observations for two precipitation events. Summing this entire period, USCRN reports slightly more precipitation (3.0) than COOP. However, the COOP station did not report precipitation for the earlier event (event 3249), instead presumably observing it all in the later event (event 3250). Because this later event was the only one with precipitation from both networks, it was the only event included in the categorical analysis of precipitation differences, which biased COOP measurements wetter by 15.8 mm in this case.
While the frequency of COOP shifts greater than 24 h is not known, one metric that might be helpful is a count of precipitating days. For the station pairs considered here, USCRN had on average reported 16 more precipitating days per station than COOP when daily totals exceeded 0.3 mm (COOP detection limit). Quantitatively, the mean network difference when both USCRN and COOP reported precipitation was much larger (3.6%) compared to overall difference (1.5%), suggesting COOP observations of precipitation tended to be aggregated over multiple days. It is important to note that this does not apply to overall or seasonal network differences when shifts in time are not as important. With that said, the impact of inconsistent reporting times on the absolute value of network differences by meteorological condition is not fully understood, but it may not have impacted trends in network differences (i.e., USCRN observing more precipitation with increasing wind speed).
This study compared two observing networks that will be used in future climate and weather studies. Using very different approaches in sensing technologies, shielding, operational procedures (manual vs automated), and QC methods, the two networks provided contrasting perspectives of daily maximum and minimum temperatures and precipitation. Temperature comparisons between stations in local pairings displayed differences that on average were similar to side-by-side comparisons, where radiative imbalances between shielding and aspiration types resulted in warmer and cooler COOP maximum and minimum temperatures, respectively.
However, in contrast to side-by-side comparisons, mean network differences for maximum and minimum temperatures were highly variable from station to station, which was unrelated to station separation or the types of sensing technology used at COOP stations (LiG or digital). Station contrasts were partially attributed to local factors including siting (station exposure), ground cover, and geographical aspects, which were not fully explored in this study. These additional factors are thought to accentuate or minimize anticipated radiative imbalances between the naturally and mechanically aspirated systems, which may have also resulted in seasonal trends. Additional analysis with more station pairs may be useful in evaluating the relative contribution of each local factor.
For precipitation, network differences also varied over monthly time scales due to the seasonality of the respective gauge biases. For instance, the unshielded COOP gauge expectedly had greater wind biases that were more pronounced over winter (less aerodynamic frozen hydrometeors) windy conditions, resulting in a dry bias with respect to USCRN. For warmer low-intensity conditions when Geonor gauge biases (wetting factor and gauge evaporation) were elevated, COOP stations had a wet bias relative to USCRN. Additional analysis of Geonor raw depths revealed that a portion of the COOP wet bias was due to gauge evaporation, which identified a possible weakness in the current USCRN algorithm used to determine precipitation from gauge depth data. Additional investigations into the performance of the USCRN precipitation algorithm and gauge sensitivity to evaporation are ongoing.
For both temperature and precipitation, COOP observer inconsistencies complicated network comparisons. Varying COOP observation times and/or multiday observations resulted in many of the larger network differences for temperature (≥20°C) and precipitation (>100%). These irregularities in addition to sensor errors (poor calibration) can result in unique bias profiles for each station that can change over time with station moves, changes in the primary observer, and instrumentation replacement. These inconsistencies complicated network comparisons and, for analysis of precipitation by prevailing meteorological conditions, may have elevated network disparities.
All observing systems have observational challenges and advantages, as noted in Table 3 for these two networks. The COOP network has many decades of observations from thousands of stations, but it lacks consistency in instrumentation type and observation time in addition to instrumentation biases. USCRN is very consistent in time and by sensor type, but as a new network it has a much shorter station record with sparsely located stations. While observational differences between these two separate networks are to be expected, it may be possible to leverage the observational advantages of both networks. The use of USCRN as a reference network (consistency check) with COOP may prove to be particularly useful in daily homogenization efforts in addition to an improved understanding of weather and climate over time as the periods of overlap between the two networks lengthen.
This work was supported by NOAA through the Cooperative Institute for Climate and Satellites—North Carolina under Cooperative Agreement NA09NES4400006. USCRN is supported by the NOAA Climate Program Office. We especially thank Scott Embler for his technical assistance; and Matt Menne, Jay Lawrimore, Tom Peterson, and the external reviewers for their editorial suggestions. The views and opinions, and findings contained in this report are those of the authors and should not be construed as an official NOAA or U.S. government position, policy, or decision.