Colorado Ongoing Basin Emissions (COBE) Estimating emissions distributions

Michael Moy1, Jenna Brown1, Arthur Santos1, Ethan Rimelman3, Olga Khaliukova2, Callan Okenberg2, William Daniels2, Dorit Hammerling2, Daniel Zimmerle1, Anna Hodshire3

1Energy Institute, Colorado State University, 2Dept. of Applied Mathematics and Statistics, Colorado School of Mines, 3Systems Engineering Dept., Colorado State University

Download Poster PDF

Background​

The Colorado Ongoing Basin Emissions (COBE) project used three companies specializing in aircraft-based methane measurements to measure emissions from a majority of oil and gas production facilities in Colorado. The companies’ technologies and measurement methods differ:

  • Bridger Photonics: Uses Gas Mapping LiDAR and cross-sectional flux estimation. Achieves a 90% probability of detection at point-source emissions of 1.27 kg/h. Bridger identifies emission sources by equipment type.
  • Insight M: Uses LeakSurveyor, a hyperspectral infrared system. Two separate sensors were used, with nominal detection limits of 10 kg/h and 25 kg/h. Insight M reports emissions at the facility level.
  • GHGSat: A shortwave infrared spectrometry system known as DATA AIR. Three separate sensors from two generations of the technology were used. GHGSat mostly reports emissions at the facility level, unless multiple plumes are clear.

Methodology​

Differences in sensor capabilities were apparent in the data:

The effect of sensor detection limits (showing sensors 1, 2, 3, and 4). Percentage of sensor's detections on the y-axis from 0% to 4%. Emission rate (kg/h) on the x-axis from 0 to 100.

Sensors are typically evaluated in controlled release tests, the results of which are summarized in (1) distributions of errors and (2) probability of detection (POD) curves, which express the probability of successfully detecting an emission as a function of the emission rate.

Insight m 10 kg/h with probability density on the y-axis from 0.0 to 0.8, relative error ratio on the x-axis from 0.0 to 3.5. Log-logistic fit α=1, β=3.432
Logistic regression for Insight m 10 kg/h sensor with probability of detection on the y-axis from 0/0 - 1.0, and emission rate (kg/h) on the x-axis from 0 to 25

Sensor errors are accounted for in our analysis by replacing each detection with a distribution estimated for the actual emission rates. POD curves let us estimate the percentage of effective samples for each sensor, as a function of emission rate. We then estimate the probability of an actual emission rate lying in an interval [a,b] as

#detections in [a,b] over # effective samples for [a,b] among all sensors

This is repeated over many intervals to estimate probability distributions summarizing the data.


Results​

Emissions seen by aircraft were summarized by probability distributions, estimated accounting for sensor POD. The examples below show the results when detections are aggregated to estimate emissions from an average facility.

Estimated distribution of facility emission rates. Two graphs with emission rate (kg/h) from 0 to 10 on the x-axis. The first chart (PDF) has probability density on the y axis going from 0./0 to 1.0 and the second chart (CDF) has the cumulative probability from 0.70 to 1.
Estimated distribution of facility emission rates
Important rare events: large emission rates with probability density on the y-axis and emission rate (kg/h) on the x-axis.
Important rare events: large emission rates

To model facilities at the equipment level, each emission was attributed to an equipment type so that distributions could be created for each equipment type. The resulting distributions were used to model facilities using METEC’s Mechanistic Air Emissions Simulator (MAES). The following example also compares the individual estimates from each of the three aerial companies.


Conclusions and Next Steps

For emission rates above 5 kg/h, the distribution can be fit well to the tail of a lognormal distribution. The method for fitting, with attention sensor capabilities, was led by the Colorado School of Mines team. Similar fits have been seen in other studies, so this range is the better-understood portion of the distribution.

Lognormal fit for facility emision rates for Bridger, Insight M, and GHGSat. On the y-axis is probability density from 0.0 to 3.0 and emission rate (kg/h) on the x-axis.
Lognormal fit for facility emision rates
Small rates: measured vs. simulated for Aerial measurements (POD-adjusted) and MAES simulation. Probability density on the y-axis from 0.00 to 0.40 and emission rate (kg/h) on the x-axis from 0 to 5.
Small rates: measured vs. simulated

Emission rates below 5 kg/h are more difficult to observe, and their importance has been recognized in recent years. The COBE measurements allow us to estimate the distribution in this range, subject to uncertainties at low emission rates. Simulations in MAES were found to produce a similar distribution of emissions in this range. Further measurements and modeling of these low emission rates are important directions for future research.


Acknowlegments and Contact Information​

Funding for COBE was provided by the Colorado Department of Public Health and Environment Agreement #2024*3364.

Michael Moy
Research Scientist
CSU Energy Institute, Colorado State University
[email protected]


References​

COBE final report:
Brown, J. A.; Moy, M.; Santos, A.; Rimelman, E.; Okenberg, C.; Daniels, W. S.; Hammerling, D. M.; Zimmerle, D.; Hodshire, A. L. Colorado Ongoing Basin Emissions (COBE) Final Report.

COBE anonymized dataset:
Brown, J. A.; Hodshire, A. Colorado Ongoing Basin Emissions Study (COBE) Anonymized Final Data Set of Emissions Measurements, 2025. https://doi.org/10.5061/dryad.8kprr4z0p.

For other COBE-related posters, see:  

Ethan Rimelman– COBE: Measurement and Source Attribution
Jenna Brown– COBE: Measurement Informed Inventory (MII) Results

Equipment at the METEC Site