Carbon monoxide (CO) is an important trace gas species due to its role as a tropospheric pollutant, its use as a tracer of atmospheric transport, and its involvement in tropospheric chemistry. The global budget of atmospheric CO involves both surface and in situ sources, as well as a single dominant atmospheric sink. Surface sources account for approximately 45 % of CO emissions and are principally composed of anthropogenically derived emissions from the incomplete combustion of fossil fuels and biofuels and emissions from biomass-burning events of both natural (lightning fires) and anthropogenic origin . In situ atmospheric CO comes from the oxidation of hydrocarbons, largely methane and isoprene, while oxidation by the hydroxyl radical (OH) is the dominant sink of CO . The reaction of CO and OH gives CO an average lifetime of 1–3 months in the troposphere and accounts for 40 % of the removal of tropospheric OH . As OH dominates the oxidizing capacity of the troposphere, the presence of CO plays an important role in modulating tropospheric chemistry. Additionally, this short lifetime is what enables CO to serve as a tracer of tropospheric transport processes as it prevents CO from becoming well mixed globally; thus, pollution sources appear as regions of significantly enhanced CO as compared to background levels e.g.,.

Due to the direct influence CO has on atmospheric chemistry, prior work has asserted that it is crucial to accurately characterize its atmospheric budget, sources, and trends e.g.,. Following industrialization, tropospheric CO concentrations increased until the early 1980s before plateauing and beginning to decrease in the 1990s and through to the present . Current estimates for global CO trends show a decrease of approximately -1 % yr-1, a decline that has been attributed to decreasing direct emissions of CO rather than changes to indirect emissions or atmospheric sinks . Trend estimates for CO are larger in the Northern Hemisphere, where most global economic activity occurs, and this decline is associated with improvements in combustion technologies that more than offset increased global fossil fuel consumption .

In contrast to the well-defined trends in CO, estimates of CO sources for use in constructing atmospheric budgets vary significantly e.g.,. A large part of this variability is due to estimates of biomass-burning emissions, which vary much more significantly than anthropogenic CO emissions from fossil fuel and biofuel consumption. These biomass-burning emission estimates typically exhibit interannual variability 2–3 times greater than that of fuel consumption, and this variability, along with the variability in indirect sources of CO, complicates efforts to model and fully understand atmospheric CO e.g.,. A major component of this is due to the variation with climate conditions, such as droughts caused by heat waves, of fire intensity and the amount of CO emitted in biomass burning events . Recent work has shown that anthropogenic climate change may lead to an increase in fire frequency and intensity, altering the global CO emission budget . This complicates the characterization of highly variable sources of CO emission, such as biomass-burning events, as climate conditions change.

The Measurements Of Pollution In The Troposphere (MOPITT) satellite instrument has been continuously monitoring CO since 2000, and it has produced the longest-running global record of CO . This data record is well suited for a variety of applications, including analysis of the variability and long-term trends in global CO distributions, examination of atmospheric transport, and exploration of the influence of human activity on global CO emissions e.g.,. MOPITT data have been used to explore emission sources of both a regular or periodic nature, such as from industry and annual cycles in anthropogenic biomass burning e.g.,, and those of an irregular episodic nature, such as the 2019–2020 Australia bushfires e.g.,. Both types of emissions need to be characterized accurately to understand atmospheric CO; the former underpins climatological signals in global CO, while the latter leads to deviations from such and can influence analysis of the long-term trends and distributions of CO .

The irregular nature of episodic emission events can make them hard to identify reliably in a dataset. Multiple methods have been employed to identify them, often employing mixes of qualitative and quantitative assays, with the most common approach involving the generation of CO anomalies by subtracting an average CO value for the period or region from a time series or climatology and then choosing a threshold for which larger/smaller values are associated with episodic events. These thresholds are often based on some multiple of the standard deviation or median average deviation of the dataset. However, this method can be ineffective if the data contain multiple extreme outliers, if there is large variability in the dataset, or if the data are multimodal or asymmetrically distributed. This paper presents a method of detecting nonseasonal emission events in the 22-year MOPITT data record using a multistep algorithm based on prior work by in detecting outlier measurements from the Atmospheric Chemistry Experiment–Fourier Transform Spectrometer (ACE-FTS) measurement dataset. Emphasis is placed within this approach on minimizing the need to directly constrain the MOPITT data, thereby generating a robust, quantitatively defined set of markers for these episodic events for use with applications such as event statistics, the analysis of individual events, and for filtering data from averaged ensembles.

The rest of the paper is organized as follows: Sect.  outlines the MOPITT instrument and the version 9 (V9) CO data products. Section  addresses the detection algorithm and the enhancement event flags produced, while Sect.  presents the results from the application of this detection algorithm. Finally, a summary is presented in Sect. .

To illustrate the results of this methodology, the measurements that are flagged as those affected by episodic events from the example time series and residual histogram shown in Figs.  and can be explored in detail. Integration of the empirically derived estimate of the EDF for this particular grid cell (centred over 30.25° S, 150.75° E; over land in New South Wales, Australia) determined the threshold value for the deseasonalized time series to be 63×1016 molec. cm-2. Given this threshold, 24 daily-mean observations are identified as statistically unlikely to have arisen, given the EDF, without the influence of an episodic CO emission event. Of these 24 observations, 7 correspond to the 2019–2020 Australia Black Summer bushfires which were some of the largest biomass-burning events on record for New South Wales . Outside of these days, five further observations are found to be coincident with other major Australian bushfires. These events include one on 23 January 2002, around the time of the Black Christmas bushfires; one on 7 November 2002, near the end of the 2002 Victoria wildfires; one on 24 January 2003, at the end of the Canberra bushfires; and a pair on 23 September 2006 and 13 January 2007 that align with the 2006–2007 Australian bushfire season. However, half of the observations flagged as those influenced by episodic emission events do not directly correspond to known major biomass-burning events in the region. Furthermore, not all known major burning events lead to observations that are flagged using this method. This underpins one of the key strengths of the detection algorithm outlined in this work, which is that it is agnostic toward prior assumptions of the impact various events can have on the MOPITT L2 data. This selective flagging allows for an extremely robust approach to data handling that maximizes the amount of information available for analysis, as it does not flag events which do not impact the CO time series in a significant manner.

Examining the global dataset, the MOPITT episodic event threshold values are shown in Fig. , while Fig.  shows the global distribution of enhancement flags for the daily-mean observations in the gridded L2 dataset. Beginning with the threshold values in Fig. , the highest threshold values are found in central South America, Indonesia, western Africa, and eastern China. Given that the seasonality and contributions from the ENSO are explicitly accounted for in the emission event algorithm, the sources of the high thresholds in these cases are frequent large CO emissions with high year-to-year variability. This leads the underlying EDFs of these regions to be very broad and results in there being few enhancement events observed where there are these high thresholds. The lack of flagging for these regions indicates that the method is functioning as it should, only highlighting daily-mean observations that are statistically unlikely to arise. This is exemplified in South America where frequent emission events are found to occur in this region, as shown in Fig. . In this example, little overlap is found between where the high threshold values are observed and where the flagged emission events occur, with the latter being located to the east of the former. This pattern also emphasizes the role of transport in enhancing the CO total column on a global basis, and it illustrates the potential for the separation of CO sources and the regions which they might impact.

From the analysis of the four regions with the highest threshold values, the source of the observed threshold values in each is found to vary. The high threshold values over central South America appear to be associated with the frequent wildfires and deforestation in the Amazon, those over Indonesia with the biomass burning during the Indonesian dry season, those over western Africa with agricultural fires in sub-Saharan Africa, and those over eastern China with widespread industrial activity and irregular annual variation in CO emissions. Within these four regions, the CO sources are associated, to at least some degree, with human activity and directly contribute to the interannual variation in these emission sources. However, these sources are captured by the broad EDFs of these regions, implying a predictable behaviour with few events that arise outside of the expected behaviour.

Figure , which shows the distribution of enhancement flags, exhibits several regions that display notably higher concentrations of episodic events than the rest of the world: namely western North America (Canada and the United States), the Amazon (Brazil), northeast Asia (Siberia), Australia, and Antarctica. The first four of these experience semi-frequent large CO-emitting wildfires that occur on a variable, nonseasonal basis. Their variable and sporadic nature identifies these events as outliers.

In contrast, the high concentration of outlier events over Antarctica is thought to arise because of CO transport. As there are few emission sources of CO in Antarctica, the CO total column over Antarctica is typically much less than the global average and very stable with little interannual variability. As a result, a relatively small quantity of CO, transported from emission sources elsewhere in the world, can greatly enhance the CO total column, leading observations to be flagged as being associated with an emission event. This property is shown in Fig. , which shows that the episodic event threshold values over Antarctica are among the lowest in the world, and small enhancements are likely to be flagged as outliers. Altogether, the combination of the enhancement flags with the event thresholds enables a condensed examination of what can be considered the typical behaviour of CO total columns on a global basis.

The temporal distribution of the flagged observations in each grid cell can also be analyzed in order to identify changes in the frequency of enhancement events over time. While the sporadic nature of these enhancement events obfuscates the detection of trends when considering the evolution in the number of these events between individual years, examining multiyear periods allows for an overview of their change with time. To this end, the episodic emission events identified, shown in Fig. , have been separated into two periods of roughly equal length, corresponding to the MOPITT measurements made between 3 March 2000 and 31 December 2010 and those made between 1 January 2011 and 31 December 2021. The difference in the number of events in each of these two periods can then be readily calculated as the number of enhancement events in latter period minus those in the former. The results of this are shown in Fig. , and from this plot several key features arise. Immediately evident is a small overall trend toward fewer enhancement events over most of the globe in the latter portion of the MOPITT measurement dataset, with the largest decreases happening over Antarctica. However, North America, Siberia, and the eastern coast of Africa all show increasing numbers of enhancement events. Addressing first the general global decrease in the number of events, while the exact cause of this general decrease is uncertain, we found that there is a high likelihood that this decrease is correlated with the decrease in global CO emissions over the MOPITT measurement period . The global decrease in CO emissions would also likely reduce the number of observed enhancement events over Antarctica, which are transport-dependent in nature and thus strongly influenced by emissions elsewhere. In contrast to this, the regions displaying elevated numbers of enhancement events in the latter period are most likely affected by an increase in fire frequency and intensity associated with anthropogenic climate change . Altogether, these findings indicate that the enhancement flags can also aid in understanding the changes in the behaviour of the CO total column over time on a global basis.

These enhancement flags can also be used to estimate the fraction of flagged observations in each grid cell associated with major enhancement events by linking flagged measurements based on their temporal distribution. For this purpose, two main factors need to be considered: the number of flagged observations required for an enhancement event to be considered a major event (clustering) and the maximum amount of time between flagged observations for them to still be considered linked (persistence). Naturally, the selection of these factors would impact the frequency and distribution of major enhancement events; however, the response to either of these factors in the number of these major events in any grid cell is influenced by the unique distribution of flagged observations in each cell.

To explore this, a set of nine classification criteria have been examined using three different values for the choice of the time frame for which observations can be considered linked and three values for the number of observations required for an event to be considered a major event. For the former, flagged observations are considered linked if they are within at most 4, 8, or 16 d of each other, with all further observations within a rolling window of that same period treated as part of the same event, while for the latter, at least two, three, or four flagged observations are required for an event to be considered major. Figure  shows the fraction of flagged daily-mean observations associated with major enhancement events for each of the criteria permutations tested. As expected, clustering generally increases with the time window permitted between flagged observations for an event and persistence decreases with the number of linked observations in the window. However, not all areas are impacted equally by each change in criteria. This property can be exploited to investigate major enhancement events by analyzing their responses to different criteria.

Here we focus briefly on five regions in Fig.  that show that a majority of their flagged observations are associated with major enhancement events under multiple classification criteria. These five regions roughly correspond to northern North America (Canada and Alaska), Siberia, the Amazon (Brazil and the western Atlantic Ocean), the east coast of equatorial Africa, and the equatorial Indian Ocean including Indonesia. Antarctica is excluded from discussion here due to the ease with which transport can dramatically enhance the CO column, as discussed above. Across these five regions, increasing clustering or decreasing persistence causes a gradual reduction in major-event fractions.

In northern North America and Siberia, there are moderate episodic event thresholds (Fig. ), fairly frequent enhancement events (Fig. ), and a reduction in the major-event fraction (Fig. ) with increasing clustering or persistence requirements for major-event classification. Together, this implies that the major events over these two regions are sporadic and likely originate within the regions themselves. As a result, it is found that increasing the stringency of the major-event classification criteria deemphasizes transport processes, restricting the area within each region displaying high major-event fractions to those areas most likely containing the direct CO emission source(s).

The Amazon shows the relationship between transport from emission sources and high major emission event fractions. In the central Amazon, high event thresholds indicate the presence of large, frequent CO emission sources. The eastward propagation of CO from these sources leads to the numerous enhancement events observed over western Brazil and the high major-event fractions seen over Brazil and the western Atlantic Ocean. Elevated episodic event thresholds and major-event fractions are also found drifting across the Atlantic Ocean to the coast of Africa, indicative of the far-reaching effects of transport from these emission sources.

The eastern coast of equatorial Africa shows significantly different behaviour than the prior regions. High episodic event thresholds, shown in Fig. , are observed over central Africa, extending westward out over the eastern Atlantic Ocean; however, there are very few enhancement events, as shown in Fig. , and very low major-event fractions over this region. This implies a very regular annual cycle in CO emissions with few deviations. The exception to the low major-event fractions occurs off the eastern coast of Africa, an area for which these fractions display high persistence and clustering. As transport from central Africa appears to be predominately westward, evidenced by the drift in high threshold values over the Atlantic, the high major-event fractions off of the eastern coast are likely the result of instances of eastward transport carrying large CO plumes out over a region with comparatively few regular emissions.

The equatorial Indian Ocean also displays significantly different patterns than the previous regions. Around Indonesia, very high episodic event thresholds are found, and few enhancement events are observed across the whole region. However, the high persistence and clustering shown by the major-event fractions, stretching from Indonesia across the Indian Ocean toward the west coast of Africa, imply the existence of occasional extreme events. Thus, the likely source of the most significant deviations in CO behaviour for this entire region stems from the westward transport of CO emissions from these extreme events.

Altogether, by combining the number of flagged observations from Fig.  with the major-event information from Fig. , a thorough exploration of the types of events identified in this study can be undertaken. From these examples, it is evident that the utility of the MOPITT enhancement flags extends beyond identifying enhancement events and into classifying groupings of these enhancements. Furthermore, grouping enhancement events to identify major events also facilitates analyses of major CO emission events on a global scale by readily identifying major events and their properties.

Motivated by the need for an improved understanding of CO, this study developed a multistep algorithm to detect days in the L2 MOPITT CO dataset that had been affected by large episodic CO emission events. This process involves deseasonalizing an observed total column time series with a centred moving average, fitting the deseasonalized time series with an empirical model of CO trends and periodic drivers, and then fitting the resulting residuals with multimodal Gaussian distributions to estimate the EDFs of the data. This latter step, which adapts the work of for detecting outliers in the ACE-FTS measurement dataset, allows for threshold values to be empirically defined, the data above which correspond to observations influenced by episodic emission events. Using these methods, the 22-year MOPITT L2 CO data record has been screened for these events, and enhancement flag files have been produced that identify corresponding days. Overall, these enhancement flags, coupled with the extensive 22-year MOPITT data record, provide insight into the distribution and frequency of these large episodic emission events and can enable more robust approaches for a wide array of applications, such as measurement validation and CO modelling, by carefully screening the data before use.