The same number does not mean the same phenomenon

If methane sensors at two farms both display 35 ppm, it is easy to assume that their emission conditions are also the same. But the first sensor may have measured for 10 seconds near a cow’s head just after feeding, while the second recorded a daily average in front of a ventilation fan. One farm may not have subtracted the outdoor background concentration, while the other may have measured the concentration of incoming air as well. Even when the displayed number and unit are identical, the subjects of comparison are different if the spaces, times, animals, and airflows represented by those numbers differ.

Measurement representativeness is a quality issue distinct from sensor accuracy. The U.S. EPA’s guidance on sampling designs for environmental data describes representativeness as the degree to which data accurately and precisely characterize a population, variation at the sampling point, and process or environmental conditions. It emphasizes that representativeness must be addressed through sampling design. A sample that is not representative cannot be repaired by improving the quality of the analytical instrument.

A statement such as “our sensor has a small error” is therefore not enough to claim a farm average or a reduction effect. A calibrated sensor may accurately read the concentration of the air that reaches it. Whether that air represents the whole farm, a particular herd, one day, a season, or the project period is a separate question. Representativeness design connects the reliability of a measurement to the scope of the claim.

First define what the value is intended to represent

Representativeness always requires a target population. The sampling unit differs depending on whether the goal is to determine the methane in one animal’s breath, the mass of methane passing through one barn outlet, or the monthly methane emissions of an entire farm. If the goal is unclear, the sensor location, averaging method, and required measurement period cannot be determined.

For livestock methane, representativeness can be divided into at least four dimensions.

  • Spatial representativeness: how well the measurement point reflects the concentration distribution within the barn and the actual paths through which air leaves.

  • Temporal representativeness: how well the measurements include feeding, rumination, milking, manure agitation, fan operation, daily temperature variation, and seasonal changes.

  • Herd representativeness: how well the sample reflects animals that differ in breed, age, weight, production stage, health, and intake.

  • Operating-state representativeness: how states such as open doors, cleaning, equipment shutdowns, hot-weather ventilation, animal placement, and shipment are included alongside normal operation.

The purpose of the analysis must also be considered. For a safety alarm, the priority is not to miss the most hazardous local or momentary concentration. For estimating farm emissions, the average and cumulative mass flow across the entire boundary matter. A test of low-methane feed must align non-feed conditions between the treatment and control groups and ensure repeatability. Do not assume that one sensor arrangement optimizes all three purposes simultaneously.

Three field situations in which accuracy and representativeness diverge

First, even if two sensors read the same concentration, different ventilation rates produce different emission amounts. The mass of methane leaving a barn generally depends on the product of the concentration increase and airflow. A reading of 20 ppm in summer when fans run quickly does not represent the same emission rate as 20 ppm in winter with less ventilation. If the outdoor backgrounds are 2 ppm and 8 ppm respectively, the barn’s contributions also differ. A concentration value describes the composition of air, not the mass emitted per hour.

Second, even the same average at the same location yields different results when the temporal samples differ. The average from 9 a.m. to 5 p.m. could be 35 ppm, and the daily average could coincidentally also be 35 ppm. The former misses nighttime rumination and changes in early-morning ventilation, while the latter includes them. Saving only the average erases this distinction. Preserve the raw time series, sampling interval, proportion of valid data, and periods with missing data together.

Third, even the same monthly average has a different interpretation when herd composition differs. The average at a farm that continuously kept 100 animals does not represent the same activity level as the average at a farm that shipped 50 animals midway through the month. Comparing only emissions per animal can also misrepresent production efficiency when milk yield, weight gain, or dry-matter intake differs greatly. Representativeness applies not only when numbers are collected, but also when denominators and normalization criteria are selected.

Does adding more sensors solve the problem automatically?

More sensors can improve representativeness, but merely increasing their number is not enough. If all five are installed near feeding points, they only repeat the same local condition with greater precision. When sensors are close together and highly correlated, the amount of information does not increase as much as expected. By contrast, stratified placement in areas with different airflows can capture important variation with fewer sensors.

The EPA’s sampling-design guidance describes selecting a design suited to the objective, such as simple random, stratified random, systematic, or grid sampling. These are not specifications that can be applied mechanically to a barn, but the principles are useful. Instead of dividing the space for convenience, stratify it by emission sources, airflow, animal density, and operating state, then obtain observations in every stratum. Mapping the conditions that undermine representativeness is more practical than searching for one “representative sensor.”

A preliminary survey is needed for this mapping. Use a portable instrument or mobile sampling line to traverse multiple heights and locations, recording wind direction, fan status, open doors, and feeding time together. Identify points with substantial spatial variation in concentration and airflow before selecting fixed-sensor locations. Survey again after seasons or ventilation modes change. There is no guarantee that a summer arrangement represents natural convection in winter.

Weights must be defined before calculating an average

Taking a simple arithmetic mean of multiple sensors is another common error. If the air passing through Outlet A accounts for 80% of the total and Outlet B for 20%, giving the two concentrations equal weight can distort the mass flow. For emission estimates, weighting tied to airflow or representative area is necessary. In naturally ventilated barns, even fixed weights may not hold because the direction of inflow and outflow at each opening changes continually.

The same applies to time averages. Do not combine 60 1-minute values and one 1-hour value with equal weight, or include a day with extensive missing data in a monthly average as if it were a complete day. Resample to a consistent time grid, establish valid-data criteria, and then examine the length and operating conditions of missing intervals. If outages were not random but occurred because equipment stopped during extreme heat, the missing data may be biased toward high-emission or high-ventilation conditions.

An average can also conceal the distribution. Of two barns with an average of 35 ppm, one may remain stable between 30 and 40 ppm, while the other stays near 5 ppm most of the time but rises to 200 ppm during manure agitation. These patterns have different meanings for cumulative carbon, worker safety, and anomaly detection. In addition to the mean, examine the median, quantiles, coefficient of variation, duration of peak values, and distributions by operating event.

How to validate representativeness

Representativeness is not demonstrated by writing “representative point” in a sensor installation plan. It must be evaluated through comparison and repetition.

The first method is simultaneous parallel measurement. Measure reference and candidate locations, and fixed sensors and mobile equipment, at the same time to examine bias and differences by operating state. Collocated measurements that place two instruments at the same location are useful for identifying instrument differences, but do not directly validate spatial representativeness. A crossover design that swaps instrument locations also helps separate instrument error from location effects.

The second method is spatial traversing and mapping. Measure concentration and air velocity at multiple points during key modes such as normal, minimum, and maximum ventilation, feeding, and cleaning. Evaluate how closely a fixed sensor agrees with the spatial mean or flow-weighted mean. If differences recur, develop a correction model or change the placement.

The third method is time-split validation. Examine how a result calculated from part of a day differs from a 24-hour or long-term reference. As the number of measurement days increases from 1 to 3, 7, and 14 days, check the independence assumptions or determine whether an uncertainty interval that reflects time-series structure—such as a block bootstrap or stratification by operating state—stabilizes. In emission estimates for naturally ventilated cattle barns, the measurement period and the number and location of measurement points can affect the result. Do not apply a minimum duration from a particular study unchanged to every farm; validate the required period using the variability of the facility in question.

The fourth method is checking against an external reference and mass balance. Compare whether the combined result from individual sensors reasonably agrees with a tracer-gas experiment, CO₂ balance, temporary chamber measurement, or independent model. Rather than treating any one method as absolute truth, examine each method’s boundary and limitations. A large discrepancy is a signal to investigate representativeness, airflow, background concentration, or time synchronization.

Create a data contract for comparability

To compare farms and periods, minimum metadata must travel with the number. Record sensor ID, model, and calibration status; coordinates and height; classification as inlet or outlet; sampling-line length; sampling interval; aggregation method; reference temperature, pressure, and humidity; outdoor background concentration; ventilation method and fan status; herd size and characteristics; feeding and manure events; missing-data rate; and substitution rules.

A data contract can also specify which claims are permitted. For example, a sensor at head height may be used for “the trend in local concentration at that location,” but not by itself for “kg CH₄ from the whole farm.” A sensor at a particular outlet can be used in a flow-weighted emission calculation while that fan is operating but excluded during natural-ventilation mode. Mark that a 7-day trial can be used for a comparison under those weather and herd conditions, but requires evidence for seasonal correction before extrapolation to an annual reduction.

Such restrictions do not diminish the value of the data. Instead, they prevent one number from being reused without limit for safety, operations, research, and carbon claims. Representativeness is not a single grade stating that data are “good” or “bad”; it is a property that defines which questions the data can answer.

Implementation checklist

  • Before comparing numbers, do the two datasets have the same target population and measurement objective?

  • Were the sensor location, height, orientation, and role relative to inlet and outlet air recorded?

  • Does the sampling schedule include variation from feeding, rumination, manure, ventilation, day and night, and seasons?

  • Were breed, weight, production stage, intake, and changes in herd size recorded?

  • Was the need for area, airflow, or time weighting examined when averaging multiple sensors?

  • Are ppm readings being compared as if they were emissions without background concentrations and ventilation rates?

  • Were instrument differences assessed by collocation and location differences assessed by spatial traversing?

  • Was it checked whether missing data are concentrated in particular weather or operating conditions?

  • Were the distribution, variability, and results by event examined in addition to the average?

  • Were permitted uses and limits on extrapolation retained as metadata for each dataset?

Conclusion: representativeness is a product of design, not a sensor property

The 35 ppm on a sensor display is information about the air that reached the sensor at that moment. It does not automatically mean that the value represents the whole farm, a month, every animal, or the outcome of a reduction project. That scope is created through a defined target population, spatial and temporal sampling design, airflow and background corrections, herd records, and validation tests.

When results differ, therefore, do not ask only which sensor is wrong. First determine what each sensor was observing, which airflows and times received more weight in the average, and whether missing data or operating changes occurred. Accurate sensors and representative measurements are both necessary, but neither replaces the other. Comparable carbon data begin by managing these two dimensions of quality separately.

Sources