Last modified on Oct 20, 2023 10:32

Contributors: Tom Gardiner (National Physical Laboratory (NPL)), David Medland (NPL), Fabio Madonna (Consiglio Nazionale delle Ricerche – Istituto di Metodologie per l'Analisi Ambientale (CNR-IMAA)), Monica Proto (CNR-IMAA), Marco Rosoldi (CNR-IMAA), Emanuele Tramutola (CNR-IMAA), Alessandro di Filippo (CNR-IMAA), Souleymane SY (CNR-IMAA), Peter Thorne (The National University of Ireland, Maynooth (NUIM))

Issued by: CNR-IMAA / Fabio Madonna

Date: 17/07/2020

Acronyms

1. Introduction

Having confidence in the observational record is key when assessing long-term changes in our climate, climate variability, and assessment of climate models and ensuing projections. Observed records being influenced by residual non‐climatic factors may lead to incorrect conclusions about the current state and evolution of the climate and, if assimilated within a reanalysis system, may lead to biases instead of positively impacting the final products. Therefore, it is important to:

-           Detect and adjust systematic inhomogeneities in the observation time series. These may arise from a variety of causes (changes in station location, instrumentation, calibration or drift issues, different instrument sensitivity across different networks, changes in measurement procedures, etc.). All such known and quantifiable effects should be adjusted for prior to application in a climate services setting;

-           Wherever possible and practicable, establish measurement traceability of ground-based observations to a reference (SI or community acknowledged) “standard” through an unbroken chain of calibrations, each contributing to the measurement uncertainty;

-           Quantify measurement uncertainties in those historical data, even where traceability cannot be properly established but instead must be estimated from the available metadata describing the measurement procedure and environment, instrumental changes, or any other influence or source of uncertainty. When metadata are partly or completely missing, analysis of sensors' uncertainty using statistical analysis of a set of measurements, or other kinds of information about the measurement process may be required;

-           Establish the comparability of results from measurements of the same variable using distinct measurement techniques, ensuring the coherent use of datasets that are physically consistent and which do not introduce any undesired effects into the analysis.

It is important to point out that it is not easy to fulfill the above requirements, especially for historical data, for global networks where much metadata is retained solely by individual station PIs (if at all) and not routinely shared or stored in the network data archives. Furthermore, for many parameters, there are a rich variety of measurement techniques that are very distinct in what, when and how they measure making it difficult to achieve full measurement comparability.

The present document provides a detailed description of the algorithms and procedures adopted by C3S to harmonize metadata and data for Near-Surface Temperature (NST) from NOAA’s U.S. Climate Reference Network (USCRN), a network of 137 stations (as of 2022) deployed across the continental U.S., Hawaii and in Alaska. The primary goal of its implementation is to provide future long-term homogeneous observations of temperature, precipitation, and soil moisture/soil temperature that can be used for current climate applications while also being coupled to past long-term observations for the detection and attribution of climate change.

2. Data and metadata sourced

The USCRN stations use high-quality instruments to measure temperature, precipitation, wind speed, soil conditions, and more. Information is available on what is measured and the USCRN station instruments (https://www.ncdc.noaa.gov/crn/).
Stations are managed and maintained by the National Oceanic and Atmospheric Administration's (NOAA) National Centers for Environmental Information.
USCRN data can be accessed via anonymous ftp at:

ftp://ftp.ncdc.noaa.gov/pub/data/uscrn/products/

and an identical web interface at:
http://www1.ncdc.noaa.gov/pub/data/uscrn/products/

In the CDS, the user may access selected subsets of monthly, daily, hourly and sub-hourly (5-minute) USCRN data. The USCRN source data archive includes measurements of air temperature, precipitation, solar radiation, surface temperature, soil moisture and soil temperature data. Calculated data are shown only if a sufficient amount of source data passes its quality control tests; otherwise, these values are set to a missing value. The C3S dataset, instead, refers only to the near-surface air temperature provided by USCRN and, in addition to NOAA's USCRN products, it provides an estimation of the total uncertainty budget also taking advantage of the other variables available within the USCRN archive. The uncertainty is calculated according to principles developed as part of GAIA-CLIM H2020 project where sources of uncertainty are identified and presented in the form of a Product Traceability and Uncertainty (PTU) document. The PTU for the USCRN is presented in Appendix A (Section 4). In Figure 1, the distribution of the USCRN stations is shown along with a picture of the station measurements setup.

Figure 1: Station distribution of U.S. Climate Reference Network (left panel) and typical station assembly (right panel).

3. USCRN documentation

The key supporting document for the CDS users is the Product Traceability and Uncertainty (PTU) for the USCRN Near-Surface Air Temperature product, provided in Appendix A. Additional documentation about USCRN is the 'Site Information Handbook', which describes the site selection process and layout, and the 'USCRN/USRCRN Data Ingest Functional Specification', which has the flowchart for which temperature measurements are accepted/rejected and whether a mean or median is used.
These can both be found at the following link: https://www.ncdc.noaa.gov/crn/documentation.html.
There is also a metadata webpage on the NOAA website at the link: https://www.ncdc.noaa.gov/isis/summary.htm?networkid=1.
Specific instruments used at the USCRN station are described on the station instruments page at: https://www.ncdc.noaa.gov/crn/instruments.html.

4. Near-Surface Temperature

The primary activity on data harmonization for NST is a metrological assessment of the overall uncertainty budget of the reference quality NST products. As part of this work sources of uncertainty in the USCRN NST product were identified. These included sources of uncertainty from:

• the instrument used,
• the data logger used,
• the interface between the instrument and the data logger and
• external effects that result from sources, such as local weather.

Detailed information on the identified sources of uncertainty is available in Appendix A, which is a PTU document for the USCRN based on the principles of assessing uncertainty developed as part of the Horizon2020 GAIA-CLIM project (http://www.gaia-clim.eu/). This also gives estimates of the uncertainty from different contributions and describes a method of combining the uncertainties to estimate the total uncertainty in the various USCRN data products covering different timescales. This uncertainty assessment is only possible because of the amount of information available about the USCRN system and the metadata collected as part of the program.

4.1. Examples of USCRN data and uncertainties

The USCRN collects temperature data using 3-sensors individually housed in fan-aspirated solar shields [1]. The USCRN [1] provides temperature measurements based on simultaneous readings from three Platinum Resistance Thermometers (PRTs). The data collected is provided to users as a 5-minute (sub-hourly) average, hourly average, daily average or monthly average. For the hourly, daily and monthly files maximum and minimums are also provided, as well as daily and monthly means which are calculated as (Tmax+Tmin)/2. Uncertainty calculations are performed using data from simultaneous measurements in the USCRN data file and using metadata made available by NOAA. A description of how the USCRN data is collected and processed is included in Appendix A.  

An example related to four USCRN sites, reported in Table 1, is shown to support users in the data full exploitation. 

Table 1: Information on sites selected for the creation of USCRN example data. 

Figures 2 and 3 provide two examples of the uncertainties estimated for the two USCRN stations considering only the temperature sensor and data logger-related contributions. In particular, Figure 2 shows the monthly average temperature for the KY-Bowling green site. The expanded uncertainty (95% confidence limit, coverage factor of k=2) in this example is consistently between ±0.1 ^oC and ±0.3 ^oC. There are two features of the uncertainty seen in Figure 8 that are related to the data-logger contribution to uncertainty. First, there is a sharp drop in uncertainty during July 2009, this is the result of replacing the data logger with a new type with a different (reduced) associated uncertainty. There are also peaks in uncertainty where the temperature falls below 0 ^oC. This occurs because the data logger has greater uncertainty in temperature ranges below 0 ^oC than above it. These effects are very noticeable because the data logger contributes such a large part to the uncertainty.
Figure 3 shows the sub-hourly average temperature and uncertainty for one year at the AK-Red Dog Mine site. Steps and peaks can be seen at 0 ^oC and -25 ^oC as a result of the data logger changing uncertainty for different temperature ranges. Another step change is clearly observed in August 2012 which is the result of how the uncertainty contribution from heating from solar radiation is currently estimated. Solar radiation measurements at this site only exist after this point and the uncertainty is therefore based on these measurements. Before August 2012, no solar radiation information is available and the uncertainty contribution is assumed to have a constant value (see Appendix A for more details).

Figure 2: Monthly average temperature and the monthly average temperature uncertainty for the USCRN KY-Bowling green site. (Date format: YYYYMM)

Figure 3: Sub-Hourly average temperature and average temperature uncertainty between Feb. 2012 and Feb 2013 at the AK-Red Dog Mine USCRN site. (Date format: YYYYMMDD)

These example plots show the complexity of the USCRN uncertainty and that various sources significantly contribute to the overall uncertainty budget. They also highlight the requirement to calculate the uncertainty on a point-by-point basis for the different data products covering different temporal periods.

Another main product of the USCRN is precipitation; incoming solar radiation and surface temperature are also recorded. These extra products are used in calculating specific parts of the uncertainty.
Because of the different ways that temperature data is provided by USCRN, and the different ways it is used, it was important to consider how the individual sources of uncertainty combine over different timescales. As a key part of this the different uncertainty sources were categorised as being systematic, quasi-systematic or random, (N.B. quasi-systematic implies that the uncertainty behaves in a consistent way but over time periods similar to that over which the measurements are recorded). A specific uncertainty source could change category depending on the timeframe over which the product was derived. For example, uncertainty from precipitation is systematic over 5 minutes, quasi-systematic over an hour and random over a day or a month. Table 2 shows the uncertainty categories for the different sources of uncertainty, and how these change over the different product timescales.

Table 2: Sources of uncertainty and their classification for the different USCRN data products.

The uncertainty of the longer time averages is calculated using information from the shorter time average data files. This is particularly true for the maximum and minimum values. An example of this is the solar radiation uncertainty of the daily average temperature, which is calculated as the average of the hourly solar radiation uncertainties. Because many of the temperatures recorded will have been at night and not affected by solar radiative heating they do not contribute to the uncertainty due to solar radiation in the daily average temperature.

The uncertainty distribution is asymmetric as some sources of uncertainty can only have a positive effect, and so could lead to a warm bias, while others can only have a negative effect due to a potential cold bias. Therefore, the uncertainty calculation provides two uncertainties, the positive uncertainty which includes any effects that could lead to a cold bias and the negative uncertainty which includes the potential warm bias effects. In the uncertainty calculation, solar radiative heating and its enhanced effect because of snow albedo are included as warming biases and cooling from precipitation is included as a cooling bias. Sensor self-heating and radiative cooling were also considered as a source of warm or cool bias respectively but their effect is much smaller than the other effects considered and is, therefore, not included in the total uncertainty estimation.

Figure 4 shows the negative uncertainty for the different USCRN time products at the OK Goodwell 2 E site, where the occasional higher values correspond to periods of snow cover, while Figure 5 shows the positive uncertainty at the same site. Because of diurnal variation in temperature and solar radiation and the frequency of precipitation the sub-hourly and hourly uncertainties appear as broad lines. The uncertainty of the average generally reduces with increasing time period down to a lower limit of about ±0.06 ^oC. This is because of the uncertainty from the datalogger, which is a symmetrical systematic uncertainty and does not reduce when averaging multiple measurements. This lower limit is higher for the earliest measurements because a different datalogger and datalogger-PRT interface was used, which has a much larger associated uncertainty.

Figure 4: Negative uncertainties for sub-hourly (blue), hourly (red), daily (yellow), and monthly (purple) average temperature for the USCRN site OK Goodwell 2 E.

Figure 5: Positive uncertainties for sub-hourly (blue), hourly (red), daily (yellow), and monthly (purple) average temperature for the USCRN site OK Goodwell 2 E.

The different files provided by USCRN and sources of uncertainty considered in the calculation give a wide range of possible outputs which can be included in a final data product The USCRN data and metadata fields available in the CDS files are described in the summary table available in Appendix B.

4.2. Uncertainties from different USCRN data products

Figure 6 and Figure 7 show the possible values for the negative and positive temperature uncertainties as well as the mean of the uncertainties for the different data products from the OK Goodwell 2 SE site. In each case, as the time length of the data product increases the mean uncertainty decreases and for data products over time scales longer than an hour the range of possible values narrows. The greatest reduction in mean uncertainty is seen between the sub-hourly and hourly data products for both the positive and negative uncertainty.

Figure 6: Negative uncertainty values and mean uncertainty for the different USCRN data products at the OK-Goodwell 2 SE site.

Figure 7: Positive uncertainty values and mean uncertainty for the different USCRN data products at the OK-Goodwell 2 SE site.


Figure 8: Possible precipitation contribution to uncertainty for different data products at the USCRN OK-Goodwell 2 SE site.


Figure 9: Possible values of the solar radiation contribution to uncertainty for different data products at the USCRN OK-Goodwell 2 SE site.

Figure 8 shows the possible values of the precipitation contribution to uncertainty for the same site. The sub-hourly uncertainty has only a limited number of possible values compared to the other data products. This uncertainty is calculated using the sub-hourly precipitation data for the hourly and higher-level data products, where there are no sub-hourly precipitation measurements the hourly uncertainty is set to its highest value. In general, the precipitation contribution to uncertainty decreases with increasing time length. This uncertainty is only present when precipitation occurs or there is no precipitation data, and as a cooling bias it is only included in the positive uncertainty. This uncertainty changes behavior twice between data products, going from sub-hourly to hourly the uncertainty changes from systematic to quasi-systematic and going from hourly to daily the uncertainty changes again to random.

Figure 10: Possible values of the snow albedo contribution to uncertainty for different data products at the USCRN OK-Goodwell 2 SE site.

Figure 11: Possible values of the PRT noise contribution to uncertainty for different data products at the USCRN OK-Goodwell 2 SE site.

There are also two warming biases included in the uncertainty calculation. The possible values that these can take for the different data products are shown in Figure 9 for the solar radiation heating and in Figure 10 for the snow albedo reflecting solar radiation. The maximum possible uncertainty reduces as the time scale of the data product increases, with the exception of the hourly data product where the maximum uncertainty occurs when there is no solar radiation measurement. The lowest possible value also increases for time scales above hourly since the mean starts to always include daylight hours, for a site at a sufficiently high latitude it is possible that the minimum uncertainty contribution from solar radiation will be 0 ^oC for some daily or monthly values. The largest change in the uncertainty contribution is seen going from the hourly data product to the daily data product, where the uncertainty changes from systematic to quasi-systematic.
In the case of the snow albedo, there is no change in how the uncertainty is determined between sub-hourly and hourly, so these show the same range of values, either the contribution is 0 ^oC when there is no snow or it is 0.4 ^oC when it has been determined that there could possibly be snow as there is no direct way of determining snow presence. This uncertainty drops off rapidly for the daily and monthly data products, where this uncertainty is quasi-systematic rather than systematic. The behaviour of this uncertainty is highly site dependent. At the OK Goodwell 2 SE site snow is fairly infrequent, but other sites can have snow events more regularly or possibly not at all.
Figure 11 shows the PRT noise contribution, which is a random uncertainty in all data products. This uncertainty has a temperature dependence and also reduces with the number of data points used in the measurements calculation. The uncertainty contribution drops by a substantial amount going from sub-hourly to hourly to daily until its contribution to the total uncertainty is very small.

Figure 12: Possible values of the datalogger contribution to uncertainty for different data products at the USCRN OK-Goodwell 2 SE site.

Figure 12 is the datalogger uncertainty. This uncertainty is systematic in all data products and the change displayed in its contribution is largely the result of the change in the measured temperature range. The uncertainty contribution from the datalogger has thresholds at 0 ^oC and 40 ^oC where the uncertainty contribution is lower between these values and increases beyond this range.

The OK Goodwell 2 SE site only uses the new type of datalogger and datalogger interface, which does not include a fixed resistor. Therefore, fixed resistor contributions are not included. The PRT calibration and lead resistance uncertainty contributions are constant and do not change between data products. The change in total positive and negative uncertainty is therefore the result of the changes in the uncertainty contributions shown in Figure 8 to Figure 12 Most of these contributions reduce as the time scale of the data product increases which means that the datalogger uncertainty, which does not reduce, quickly becomes the most influential term in the total uncertainty with increasing timescale. Figure 13 shows how the total uncertainties evolve over different timescales for the OK Goodwell 2 SE site.

Figure 13: Possible (points) and mean (solid line) uncertainty values for different data products from the OK-Goodwell 2 SE site. The positive and negative uncertainties are shown in the left and right panels respectively.

Data access and data format

The data are available and described in the "Overview" page of the Climate Data Store (CDS)<link>.

The CDS web interface provides data in two CSV formats:

• Level-wise rows;
• Observation-wise rows. The option to download the data NetCDF4 format will become soon available. All CDS in-situ observations share a common data model (CDM-Obs). This format is described at https://glamod.github.io/cdm-obs-documentation/introduction.html.

4.3. Dataset tabular description

Table 3: Subhourly data

Table 4: Hourly data

Table 5: Daily data

Table 6: Monthly data

5. Product Availability and data license

The GRUAN data policy can be found in the "License" section of the related "Overview" page in the CDS.

6. Acknowledgements

The Authors would like to thank Dr. Andrea Merlone at the Istituto Nazionale di Ricerca Metrologica (INRiM) for contributions to the method for calculating the snow albedo and precipitation uncertainties and Howard Diamond at the National Oceanic and Atmospheric Administration for correspondence on the details of the USCRN facilities.

7. References

1 Diamond, H. J., T. R. Karl, M. A. Palecki, C. B. Baker, J. E. Bell, R. D. Leeper, D. R. Easterling, J. H. Lawrimore, T. P. Meyers, M. R. Helfert, G. Goodge, and P. W. Thorne, 2013: U.S. Climate Reference Network after one decade of operations: status and assessment. Bull. Amer. Meteor. Soc., 94, 489-498. doi: 10.1175/BAMS-D-12-00170.1

8. Appendix A: Product Traceability and Uncertainty for the USCRN Near-Surface Air Temperature product

Version 2.3




Date: October 2018





Compiled by David Medland and Tom Gardiner (NPL)

8.1. Introduction

This document describes the product traceability and uncertainty of the USCRN (United States Reference Climate Network) air temperature data product. This is done according to principles developed as part of the GAIA-CLIM H2020 project (http://www.gaia-clim.eu/). The USCRN is a network of 137 stations collecting data on a number of variables with the aim of providing long-term climate observation. Work on the USCRN began in June 2000 with the first sites starting recording in January 2004. Data collection and processing are carried out by the NOAA (National Oceanographic and Atmospheric Administration) and processed data is available at https://www.ncdc.noaa.gov/crn/data.html.

Following this introduction, a summary is provided of the near-surface temperature measurements made by the USCRN. The traceability chain for these measurements is shown in Section 4, which identifies all of the individual components of the overall measurement uncertainty. These are then described in Sections 4.5 to 4.25, and summarised in Section 4.26.

8.2. Near-Surface Air Temperature Measurement

Air temperature is one of the primary measurements of the USCRN and sites are expected to achieve a minimum accuracy of ±0.3 ^oC over the range -50 to +50 ^oC and ±0.6 ^oC outside this range, to -60 and + 60 °C. Sites are chosen to be representative and minimize local effects. Each site uses 3 platinum resistance thermometers (PRTs) inside their own aspirated solar shield held 1.5 m above the ground. The thermometers used are Thermometrics PT1000 PRTs. The solar shields used are Met One Instruments model 076B 7308, modified for easier maintenance. The solar shield acts to reduce the heating of the thermometer by solar radiation during the day and cooling by outgoing infrared during the night. Temperature results are recorded using a datalogger, originally a Campbell Scientific, Inc. CR23X datalogger, later replaced with a Campbell Scientific CR3000 micrologger.

The sensors record temperature every 10 seconds, these measurements are then recorded as a 5 minute mean. If all 3 sensors are within 0.3 ^oC of each other and all 3 fans are working at optimal speed then the median is chosen as the reading for that 5-minute period. The use of median values acts to reduce the random uncertainty of the measurements, but the scale of the reduction is less than the √3 (1.73) reduction that be achieved by taking the mean of the 3 readings (however the result is less sensitive to individual outliers). A simple Monte-Carlo model of the process shows that the effect of resampling a normal distribution using the median of three samples gives an uncertainty reduction factor of 1.47. If only two sensors agree or only two fans are working then the average of the two sensors is recorded. The hourly and daily averages, extremes and means are found from these 5-minute means. The monthly average, maximum and minimum are the averages of the daily values.

The USCRN reports sub-hourly, hourly, daily and monthly datasets. The sub-hourly contains the 5-minute average temperature. The hourly, daily and monthly include the average over the time period as well as maximum and minimums. The daily and monthly also include a mean value, which is calculated as (maximum+minimum)/2, compared to the average which is calculated from all available measurements.

Sources of uncertainty come from the equipment used in the measurement, including the sensor, the datalogger and the sensor-datalogger interface, and outside influencing effects, such as solar radiation. These sources contribute a range of random and systematic components to the overall uncertainty. Some uncertainties will not change over the time period for which different measurement products can be derived, e.g. the bias from solar heating will generally not change over the hourly means, but will over daily and monthly which include night-time measurements where the bias is not present. There would still be some uncertainty in the longer timescale product as a result of these biases but it will be reduced by the measurements where the bias is not present. Such uncertainties are identified as a 'quasi-systematic'. Some systematic uncertainties are present in all measurements and so will not change or reduce with data averaging.

8.3. Traceability chain

Figure 4.1: Traceability chain for the USCRN near-surface temperature data product

8.4. Element Contributions

8.5. 1, PT1000, U_PRT

The thermometer used by the USCRN system is a Thermoemtrics PT1000 which is a Platinum Resistance Thermometer (PRT) that has a resistance of 1000 Ω at 0 ^oC. The resistance of the PRT changes with temperature allowing a datalogger, see section 4, to calculate the temperature of the PRT based on the voltage across it. Information on the sensor can be found at: https://www1.ncdc.noaa.gov/pub/data/uscrn/documentation/site/sensors/airtemperature/Descriptions/AirTemp_ThermometricsPT1000PRT.pdf

8.6. 1a, PT1000 measurement Noise, U_acc

The measurement uncertainty of the PT1000 is given by USCRN in terms of the resistance of the PRT, the accuracy in temperature will depend on the calibration equation of the specific sensor. Here, it has been given based on the coefficients for one sensor at the TX-panther junction site. How this average will depend on the type of uncertainty it represents. It is assumed that the reported uncertainty represents the individual measurement variability and is random from measurement to measurement and sensor to sensor.

8.7. 1b, NIST traceable calibration

During the annual maintenance of each site, one of the 3 PRTs is replaced with a new calibrated PRT, so that each sensor at a site is calibrated every 3 years. The calibration equations take the form T = c₀ + c₁R + c₂R² where R is the resistance of the PRT and c_i are coefficients determined by the calibration process. These coefficients can be found for each sensor in the metadata tables for the site, available online at https://www.ncdc.noaa.gov/isis/stationlist?networkid=1. Uncertainty will be systematic over time for a single sensor.

8.8. 2, USCRN PRT assembly, U_local

The PRTs are held inside an aspirated solar shield such that air is sampled 1.5 m from the ground. The situation in which the measurement is taken introduces a number of biases, such as from the current passing through the PRT and the effect of solar radiation, etc. The uncertainties which contribute to this section are the result of biases that result from the sensor or USCRN system design or site location. These would not be expected to average out between the sensors. Some effects are the result of short-term phenomena and so will not have an effect on every measurement.

8.9. 2a, Self-heating, δT_SH

Self-heating error occurs as a result of the power of the current passing through the PTR. The self-heating power of the USCRN PRT is about 0.55 mW. For standard PRTs self-heating error is usually below 0.001 k, increasing with temperature but very low in the temperature regions that the USCRN system is expected to operate. The result is a warm bias from the sensor design that will not average out across multiple measurements. However, its small magnitude will mean its contribution to the final uncertainty will be minimal.

8.10. 2b, Site surroundings, δT_site

The USCRN site selection process uses a classification system to determine the representativeness of the site. Large errors are associated with sites classed 3-5 as a result of the proximity to heating sources, this can be found in the USCRN site specification guide.
Class 1 – Flat and horizontal ground surrounded by a clear surface with a slope below 1/3 (<19º). Grass/low vegetation ground cover <10 centimeters high. Sensors located at least 100 meters from artificial heating or reflecting surfaces, such as buildings, concrete surfaces, and parking lots. Far from large bodies of water, except if it is representative of the area, and then located at least 100 meters away. No shading when the sun elevation is >3 degrees.
Class 2 – Same as Class 1 with the following differences. Surrounding Vegetation <25 centimeters. Artificial heating sources within 30m. No shading for a sun elevation >5º.
Class 3 (error 1ºC) – Same as Class 2, except no artificial heating sources within 10 meters.
Class 4 (error ≥ 2ºC) – Artificial heating sources <10 meters.
Class 5 (error ≥ 5ºC) – Temperature sensor located next to/above an artificial heating source, such as a building, rooftop, parking lot, or concrete surface.
USCRN sites are chosen to reduce local temperature effects, and it is beyond the scope of this work to carry out a site-by-site assessment. Therefore, the uncertainty contribution from this effect is currently not included in the overall measurement uncertainty.

8.11. 2c, Outgoing thermal radiation, δT_TR

The PRT sensor will lose heat as outgoing thermal radiation, producing a cooling bias. Thermal radiation is determined by the Stefan-Boltzmann law, j=AεσT⁴, where A is the surface area, ε is emissivity, σ is the Stefan Boltzmann constant, T is the temperature in kelvin and P is the power emitted across all wavelengths. Platinum has an emissivity of 0.05 at the temperatures seen by the USCRN, assuming a surface area of around 1.5*10^-5 m² then the thermal radiation emitted by the PRT has a power of about 0.32 mW at a temperature of 290 K/ 17 ^oC. This is smaller than the self-heating power so would be expected to produce a smaller error.

8.12. 2d, precipitation, δT_Precip

Precipitation can lead to a cooling bias, even for a shielded and aspirated thermometer. This uncertainty results from the difference between rain and NST temperature. The value for the uncertainty used here has been determined based on a sample of measurements from an experiment investigating the difference in measured temperature between a temperature sensor simulated rainfall and one in dry conditions. This experiment included sensors in aspirated shields, although of a different design to USCRN. The uncertainty was seen to vary with the difference between air and precipitation temperature and here we have used a value for uncertainty assuming a rain-air temperature difference of 2.0^oC based on the results of Byers et al. (1949) from multiple precipitation events.

The maximum uncertainty of 0.2^oC is applied on the first data point in a precipitation event and then decays linearly over 30 minutes to 0.05^oC. This is to represent the reduction in precipitation/air temperature difference as the air cools. The uncertainty then stays at 0.05^oC until 30 minutes after the precipitation event has ended. The uncertainty continues because of differences in temperature that may result from evaporative cooling. The maximum precipitation uncertainty is not applied again until at least 30 minutes of no precipitation have passed, if there is a break in precipitation of fewer than 30 minutes it is considered a single precipitation event and the uncertainty of 0.05^oC continues. The timescales for the precipitation effects are based on the most common scenario seen by Byers et al. Since the sub-hourly data is given at 5-minute intervals the precipitation uncertainty occurs over 6 fixed levels spaced evenly from 0.2^oC until it falls to 0.05^oC, 30 minutes after the start of the precipitation event. For the USCRN data products, for time periods longer than an hour the uncertainty is calculated based on the uncertainty of the sub-hourly data product, see section 7, uncertainty combinations.

Because this uncertainty is the result of a cooling bias the contribution it has to the total uncertainty is asymmetrical. Therefore, we have separated the total uncertainty into two parts, the positive uncertainty containing the cooling biases and the negative uncertainty containing the warm biases. If these were plotted as error bars, the positive uncertainty would be the top error bar and the negative uncertainty the bottom error bar.

8.13. 2e, Sensor radiation error, δT_SR

Incoming solar radiation can land on the PRT, heating it and causing a warm bias. Investigations into measurements in non-aspirated solar shields on buoys have found that errors may reach as high as 3^oC, but the USCRN system is designed to greatly reduce this using an aspirated solar shield. Solar radiation is measured as part of the USCRN program using a Kipp and Zonen Inc SP Lite2 silicon pyranometer, allowing some attempts at a more accurate estimate of uncertainty from measurement. Would influence daytime hourly and sub-hourly means but the uncertainty would be reduced in the daily and monthly means. The maximum value for this effect is estimated to be 0.05^oC, so a simple uncertainty model has been introduced that weights this maximum uncertainty by the ratio of the measured solar radiation to a nominal maximum value of 1600 Wm^-2. Early USCRN data does not include solar radiation, so the maximum uncertainty is currently assumed where no data is available.

8.14. 2e1, Aspirated solar shield

The USCRN uses Met One Instruments model 076B 7308 aspirated solar shields that have been modified for easier maintenance. These reduce the influence that incoming radiation and precipitation and outgoing radiation have on the temperature of the PRT. The effectiveness of the solar shield can decrease over time, but this effect in the USCRN system should be reduced by aspiration. The USCRN instrument description can be found at: https://www1.ncdc.noaa.gov/pub/data/uscrn/documentation/site/sensors/aspiratedshield/Descriptions/AirTemp_MetOneAspiratedShield.pdf.

The fan speed is monitored and if a fan slows below 80 % of the optimal speed the result from the sensor is discounted.

It is assumed that the effect of the solar shield is included in the solar radiation uncertainty contribution (2e).

8.15. 2e2, Surface albedo

The surface can reflect solar radiation onto the PRT, particularly if it is snow-covered. Soil and grass have an albedo of about 0.2 to 0.25 while snow can have an albedo 0f 0.8 to 0.9. The increase in upward solar radiation can have a large effect on even fan-aspirated solar shields. This bias would only appear if there is sufficient sunlight after the snow has settled. It is not recorded by USCRN if there is snow on the ground but we make a prediction using the IR surface temperature measurements, precipitation measurements, and solar radiation measurements. The current error estimation includes the contribution from the reflected solar radiation when the following conditions are met:

1. The surface temperature is 0^oC, if the surface temperature is above 0^oC then there can be no snow. In the absence of surface temperature measurements, the condition is instead to check if NST is below 2^oC.
2. There has been precipitation equivalent to 1 mm liquid precipitation, which should be equivalent to 10 mm solid precipitation. If concurrent precipitation measurements are missing then it is assumed that this condition has been met.
3. There is recorded solar radiation. If there are no concurrent solar radiation measurements then the time of the temperature measurement is checked against sunrise and sunset times using the calculation found at https://www.esrl.noaa.gov/gmd/grad/solcalc/calcdetails.html.

The uncertainty has been determined based on studies of temperature measurements in aspirated solar shields by INRiM. There, a difference of 0.4^oC was observed between a measurement over snow and one taken simultaneously by the same set-up over an area where the snow had been cleared. There was only a very slight variation of the difference with reflected solar radiation and so in this uncertainty estimation it is assumed to be constant as long as there is incoming solar radiation to be reflected.

8.16. 3, Sensor-datalogger interface, U_int

The USCRN system used two methods of connecting the PRT to the datalogger, depending on the datalogger used.
For the CR23X the PRT sensor is connected to the datalogger by a 3-wire-half bridge containing a 1000 Ω fixed resistor. The uncertainty contribution of the sensor-datalogger interface is a combination of uncertainty as a result of the resistance in the connecting leads and the uncertainty in the resistance of the fixed resistor. The USCRN system originally used a 1000 Ω fixed resistor with a 1500 mV voltage across it with a 2000mV full-scale range, in 2006 this was changed to a 10000 Ω fixed resistor with 1800 mV across it and a 400 mV full-scale range. Current uncertainty processing is for post-2006 measurements and so uses these values.
For the CR3000 the PRT is connected directly to the datalogger and uses an excitation current of 167 μA. For this interface, there is no fixed resistor and so the interface uncertainty is only from the leads used.

8.17. 3a, Fixed resistor error, U_res

The datalogger measures the voltage across the PRT as a ratio of the voltage across a fixed resistor. The accuracy to which the voltage across the fixed resistor is known depends on the accuracy to which the resistance of the fixed resistor is known. There are two parts to this, the tolerance of the fixed resistor and how the resistance of the fixed resistor can be expected to change with temperature.

8.18. 3a1, Fixed resistor tolerance, U_tol

The fixed resistor, R_ref, had an associated uncertainty in its stated resistance of 0.01 %. Each sensor has a separate fixed resistor so the uncertainty will reduce when the combination of measurements is taken across sensors, but for each sensor the uncertainty will be consistent across measurements so the uncertainty will not reduce when the mean over time is calculated.

8.19. 3a2, Fixed resistor temperature coefficient, U_Tco

The fixed resistor resistance varies from its stated value with temperature. Here the uncertainty is represented by how much the resistance can be expected to change with a temperature change from 25^oC.

8.20. 3b, Lead resistance, U_lead

The leads connecting the sensor and fixed resistor to the data logger will also themselves have resistance and, if this resistance is different between leads, it will affect the measured ratio of voltage across the PRT to the reference voltage.

8.21. 4, Datalogger

Two models of datalogger are used by the USCRN through different time periods of operation. These are the Campbell Scientific CR23X and CR3000. The uncertainty in ratiometric voltage readings by the CR23X is:

where FSR is ±1000mv. The CR3000 ratiometric measurement uncertainty is:

Where the offset in this case is 51.1 μV. The specification information can be found at: https://s.campbellsci.com/documents/us/product-brochures/s_cr23x.pdf
and https://s.campbellsci.com/documents/us/product-brochures/s_cr3000.pdf
Because the same datalogger is used for all 3 sensors at a site and is consistent across measurements the uncertainty will not reduce with the combination of sensors or mean over time.

8.22. 5, 5 minute mean

The data distributed by USCRN is not the raw data provided by the instruments. Three sets of PRTs, each recording temperature every 10 seconds, are used and data from each sensor is stored on the datalogger in 5 minute means. If the fans on all 3 aspirated shields are working at an acceptable speed and all sensors are within 0.3^oC of each other the median is chosen as the value for that time. If only 2 sensors are within 0.3^oC or one fan is working two slow the mean of the two sensors is chosen. If no sensors agree or two or more fans are too slow no measurement is recorded and NaN is reported instead. Taking the mean across the 5 minutes will reduce uncertainties which are random between measurements and taking the median or mean of the sensors will reduce uncertainties which are random between sensors. Systematic effects which work over timescales longer than 5 minutes, such as solar radiation and precipitation effects, will not be reduced. In 98 % of cases the value is the median of the 3 sensors.

8.23. 6, hourly mean, hourly maximum/minimum, T_calc.

The hourly mean is the mean of the 5-minute values calculated previously. The hourly minimum and maximum is the median of the lowest and highest 5-minute means from each sensor, these will not necessarily be from the same time. The uncertainty on these values will be the same as calculated in the 5-minute mean section. T_calc is the mean of the last 5 minutes of the hour and so will also have the same uncertainty as calculated in the previous step.

8.24. 7, Daily mean, average, maximum and minimum

The daily average is calculated from the 5-minute means similar to the hourly average, this reduces the uncertainties which are random between measurements and also the uncertainties from effects that last less than a day, for example solar radiation.
The maximum and minimum are the median of the highest/lowest 5-minute mean from each sensor that day, and so the uncertainty will be the same as calculated in section 4.5. The daily mean is calculated from the (maximum+minimum)/2.

8.25. 8, Monthly mean, average, maximum and minimum

The monthly values for average, maximum and minimum are calculated as means of the daily values. This will reduce random uncertainties and systematic uncertainties that work over timescales of less than a month. The monthly mean is calculated as (monthly maximum + monthly minimum)/2.

8.26. Uncertainty Summary

Most uncertainties included are small, the largest potential sources of uncertainties are the heating from surface-reflected solar radiation and precipitation, although these uncertainties do not apply to all measurements. The site uncertainty may be much larger, but this should be minimised by the site selection process. The largest uncertainty that is applicable to all measurements is the datalogger uncertainty. The next most significant uncertainty would be the sensor-datalogger interface and the sensor accuracy. After this, the uncertainty contributions are small, on the order of 1/100 to 1/1000 of a degree. Most of the uncertainty contributions are systematic, but depending on the temperature and variability over the relevant time period the uncertainty may arise mostly from the random uncertainty contributions. The USCRN program aims to keep uncertainty below 0.3^oC however there are conditions when the uncertainty would be expected to be higher, particularly at the extremes of the USCRN operating ranges where the datalogger uncertainty is high.

Some uncertainties are included in the final product but are hard to estimate. The uncertainty contributions from precipitation and surface-reflected solar radiation, which are the largest sources of uncertainty in some measurements, will vary heavily depending on the system design. While we have used estimates based on temperature sensors in fan-aspirated solar shields it is not clear if a system like USCRN specifically will behave in exactly the same way.

8.27. Uncertainty Combinations

USCRN provides data over several time steps that is calculated from multiple measurements. Different uncertainties may be reduced depending on the source of the uncertainty.

• Random – the error changes between measurements.
• Systematic – the error behaves in a consistent way between measurements over a time period longer than the time period over which measurements are recorded.
• Quasi-Systemic – the error behaves in a consistent way but over time periods similar to that over which the measurements are recorded.

The different sources of uncertainty have been classified according to these criteria as shown in table 1. In the uncertainty estimation, uncertainties that change from systematic in one time period to quasi-systematic or random in the longer time periods are calculated from the shorter period time products, for example in the sub-hourly average temperature the uncertainty from precipitation is calculated as described in section 5 but in the hourly it is calculated as the mean of the precipitation uncertainties of the sub-hourly averages contained within the same time period. When a source of uncertainty changes from quasi-systematic to random, this mean is reduced according to the number of data points used to calculate it. Although USCRN does not provide a long-term multiple month data product how the uncertainties behavior is expected to change has been provided in Table 4.1.

Table 4.1: Sources of uncertainty and their classification for the different USCRN data products.

8.28. Maximums, Minimums and Means

The hourly, daily and Monthly files contain maximum and minimum information for those time periods. The uncertainty for the maximum and minimums is calculated separately from the averages. For the daily and hourly files, the maximum and minimums depend only on measurements over a 5-minute period. The process by which the hourly maximum and minimums are chosen are described in the USCRN data ingest function document while the daily hourly maximum is the highest hourly maximum and the daily minimum is the lowest hourly minimum for that day. The uncertainty of the hourly maximum and minimum is calculated in a similar way to the sub-hourly uncertainties, using worst-case solar radiation and precipitation uncertainties for that hour in cases where the time the maximum or minimum were measure is unknown. The daily maximum and minimums are taken from the hourly data processing unmodified. The hourly and daily means uncertainty contributions are classified in the same way as for the averages and their uncertainties are there for calculated in a similar manner, except with fewer data points included in the mean. The monthly maximums and minimums are the means of the daily maximums and minimums and the uncertainty contributions are classified in the same way as the monthly averages, and the uncertainty calculations is done in the same way as for the monthly average, except using information about the daily maximum and minimum uncertainties rather than the daily average uncertainty. Because these are based on fewer data points, and because of assumptions made in the hourly case, the uncertainties for maximums, minimums and means are generally higher than the uncertainties for their corresponding averages.

8.29. Example Uncertainty

The combined positive and negative uncertainties for the KY-Bowling Green hourly data are shown as a time series in Figure 4.2. This location (at 37.2503^o latitude, -86.2325^o longitude) has been randomly selected to provide a case study of the typical results of the uncertainty analysis. These demonstrate a number of features of the uncertainty behavior:

• The positive uncertainty looks like it is made of more lines than the negative uncertainty, this is because precipitation events are frequent at this site.
• The negative uncertainty has infrequent periods where it is much larger than the positive uncertainty, this is where there are occasional contributions from the snow albedo uncertainty.
• There is a sudden reduction in uncertainty in 2009, this is when the CR23X datalogger has been replaced by the CR3000.

Another change in the behavior of uncertainty with time is shown in Figure 4.3, which shows the contribution to the uncertainty by solar radiation error for the KY-Bowling Green sub-hourly data. Before around September 2012, there is no recorded solar radiation data so an uncertainty is applied according to whether the measurement is determined to have taken place between sunrise and sunset times. This is why it appears as two lines, one at 0 uncertainty when the measurement is at night and one at 0.05 when the measurement is during the day. With concurrent solar radiation measurements, the contribution to the total uncertainty is generally smaller and more varied.

Figure 4.2: The combined negative (blue) and combined positive (red) uncertainties for the sub-hourly data KY-Bowling Green site.

Figure 4.3: The contribution of solar radiation to the temperature uncertainty of the sub-hourly data for the KY-Bowling Green site. There is a line at 0.05 and at 0 because uncertainty is applied according to the time of measurement and local sunrise and sunset when there is no solar radiation data available.

Figure 4.4 shows the uncertainty against temperature for the KY-Bowling Green site sub-hourly data for the contributions from the PRT, PRT-datalogger interface and the datalogger. These are all either constant or have a temperature dependence, compared to the solar radiation, snow albedo and precipitation uncertainties which are not constant but do not have a direct temperature dependence. The greatest of these is the datalogger uncertainty, which appears at two different levels because of the two dataloggers used. This also shows the change in datalogger uncertainty for temperatures above and below 0^oC. The other uncertainty contributions are fairly similar in magnitude. Because of how it is calculated the resistor temperature dependence appears to go negative but because uncertainties are combined in quadrature its inclusion will always increase uncertainty, not decrease it. The fixed resistor tolerance and temperature dependence both have two lines, with one going along the y=0 line, because they do not contribute to the uncertainty when the CR3000 datalogger is used.

Figure 4.4: Different uncertainty contributions against temperature from the sub-hourly data for the KY-Bowling Green site., the datalogger uncertainty appears to multiple two lines because of how uncertainty is calculated above and below 0^oC and because of the two data loggers used. The higher uncertainty is associated with the CR23X and the lower with the CR3000. The fixed resistor uncertainties both have lines going along y=0 because they do not contribute to the uncertainty of measurements recorded by the CR3000.

The total combined positive and negative uncertainties are shown against temperature for the KY-Bowling Green sub-hourly data in Figures 4.5 and 4.6 respectively. These both appear as multiple levels, partly because of the two different dataloggers used and partly because the solar radiation, snow albedo and precipitation do not contribute to the uncertainty of every measurement. The highest negative uncertainties can be seen when there is a snow-albedo contribution, which is only at lower temperatures. The positive uncertainty has multiple different levels because of how the precipitation uncertainty is applied, as one of 6 possible values (excluding 0).
For all temperature measurements, the uncertainty is above 0.06^oC. With the exception of some measurements during precipitation events or with snow presence uncertainty is below 0.15^oC. With the newer datalogger the uncertainty is usually below 0.1^oC.

Figure 4.5: The combined negative uncertainties against temperature from the sub-hourly data for the KY-Bowling Green site.

Figure 4.6: The combined positive uncertainties against temperature from the sub-hourly data for the KY-Bowling Green site.

8.30. Further Examples

Figures 4.7 to 4.9 show the uncertainty of sub-hourly measurements over time for 3 USCRN sites, in the same manner as Figure does for KY Bowling Green. Figure 1.7 shows AK Red Dog Mine, Figure shows OK Goodwell (the 2 E site, there is also an OK Goodwell 2 SE) and Figure 4.9 shows HI Hilo.
The OK and HI sites both used the CR23X datalogger for the earlier measurements while the AK site only used the CR3000.

Figure 4.7: Negative (blue) and positive (red) uncertainty for sub-hourly temperature measurements from the USCRN site AK Red Dog Mine 3 SSW.

Figure 4.8: Negative (blue) and positive (red) uncertainty for sub-hourly temperature measurements from the USCRN site OK Goodwell 2 E.

Figure 4.9: Negative (blue) and positive (red) uncertainty for sub-hourly temperature measurements from the USCRN site HI Hilo 5 S.

8.31. Recommendations

The uncertainties from reflected solar radiation by snow and from precipitation have been determined using studies of systems that are different from the USCRN system. These are also the largest uncertainty contributors to the positive and negative uncertainty, although they do not contribute to every measurement. Studies to determine these uncertainties using the actual USCRN system would produce more accurate estimations which could possibly be much lower than those used in the current uncertainty calculation. Snow presence on the ground is determined from proxy data that only indicates if it is possible that snow is present. Some methods of determining if there is snow present, such as a downward-facing solar radiation detector to look for reflected light, would mean that this uncertainty is more accurately applied.
USCRN does provide more information about the individual element contributions in the form of 5-minute means broken down into separate temperature sensors, this could be useful in better determining the uncertainties which are based on medians or in more accurately determining the number of measurements from which a mean is calculated. The method through which this information is made available is however fairly user-unfriendly and this information was not involved in the data processing.

8.32. References

NOAA, USCRN Website, https://www.ncdc.noaa.gov/crn/overview.html, accessed 14/07/2018

Diamond, H. J., T. R. Karl, M. A. Palecki, C. B. Baker, J. E. Bell, R. D. Leeper, D. R. Easterling, J. H. Lawrimore, T. P. Meyers, M. R. Helfert, G. Goodge, and P. W. Thorne, 2013: U.S. Climate Reference Network after one decade of operations: status and assessment. Bull. Amer. Meteor. Soc., 94, 489-498. doi: 10.1175/BAMS-D-12-00170.1

NOAA-NESDIS, United States Climate Reference Network (USCRN), FY 2006 Annual Report, https://www1.ncdc.noaa.gov/pub/data/uscrn/publications/annual_reports/FY06_USCRN_Annual_Report.pdf

BIPM, Guide to the realisation of the ITS-90, Section 5 ‘Experimental Sources of Uncertainty’, 2018, https://www.bipm.org/utils/common/pdf/ITS-90/Guide_ITS-90_5_SPRT_2018.pdf

INRIM(a), Private communication on the increase in uncertainty of temperature sensors from precipitation. 30/07/2018.

Byers, H. R., Moses, H., Harney, P. J., 1949, Measurement of rain Temperature, Journal of Meteorology.

INRIM(b), private communication on the increase in temperature albedo from reflected solar radiation as a result of snow. 30/07/2018.

Lin, X. and K.G. Hubbard, 2004: Sensor and Electronic Biases/Errors in Air Temperature Measurements in Common Weather Station Networks. J. Atmos. Oceanic Technol., 21, 1025–1032, https://doi.org/10.1175/1520-0426(2004)021<1025:SAEEIA>2.0.CO;2

NOAA, Climate Reference Network, Site Information Handbook, 2002, https://www1.ncdc.noaa.gov/pub/data/uscrn/documentation/program/X030FullDocumentD0.pdf

Hubbard, K.G., X. Lin, and C.B. Baker, 2005: On the USCRN Temperature System. J. Atmos. Oceanic Technol., 22, 1095–1100, https://doi.org/10.1175/JTECH1715.1

USCRN, Private communications on the PRT-Datalogger interface, 20/09/2018

NOAA, USCRN/USRCRN Data Ingest Functional Specification, 2013, https://www1.ncdc.noaa.gov/pub/data/uscrn/documentation/program/CRN_Ingest_Functional_Spec.pdf.

Related articles

• Page:
  Alpine gridded monthly precipitation data since 1871 derived from in-situ observations (LAPrec): user guide and technical report
• Page:
  E-OBS daily gridded observations for Europe from 1950 to present: Product user guide
• Page:
  Nordic Gridded Climate Dataset (NGCD): Known issues
• Page:
  Nordic Gridded Climate Dataset (NGCD): Algorithm Theoretical Basis Document (ATBD)
• Page:
  Climate indices derived from E-OBS: Product User Guide (PUG)