Trends and Interannual Variability of the Hydroxyl Radical in the Remote Tropics During Boreal Autumn Inferred From Satellite Proxy Data

Despite its importance for the global oxidative capacity, spatially resolved trends and variability of the hydroxyl radical (OH) are poorly constrained. We demonstrate the utility of a tropospheric column OH (TCOH) product, created from machine learning and satellite proxy data, in determining the spatial variability in trends of tropical OH over the oceans during September through November. While OH increases domain‐wide by 2.1%/decade from 2005–2019, we find significant spatial heterogeneity in regional trends, with decreases in some areas of 2.5%/decade. Our analysis of the trends in the proxy data indicate anthropogenic‐driven changes in emissions of OH drivers as well as increasing temperatures cause these trends. This OH product is potentially a significant advance in constraining OH spatial variability and serves as a useful complement to existing tools in understanding the atmospheric oxidative capacity. Comprehensive observations of TCOH are required to assess the fidelity of this method.


Introduction
The hydroxyl radical (OH) is the primary atmospheric oxidant, setting the photochemical lifetime of many tropospheric constituents, including methane (CH 4 ).Despite this importance, OH distributions, trends, and variability are not well-constrained with observations, leading to large uncertainties in the CH 4 budget (Saunois et al., 2020).Hydroxyl's short lifetime (Stone et al., 2012) and low abundance (Mao et al., 2009) lead to large spatiotemporal heterogeneity that a sparse, surface monitoring network cannot capture.In situ observations are therefore generally limited to intense field campaigns (Miller & Brune, 2022) and are unsuitable for constraining long-term OH variability.Satellite-based OH observations are confined to the stratosphere (Pickett et al., 2008).
Because of these sampling difficulties, trends and variability in tropospheric OH are generally inferred from other trace gases, such as methyl chloroform (MCF), with well-characterized emissions and OH as their only or primary sink.Inversions using box models and MCF surface observations yield information about OH variability but only at global or hemispheric scales (e.g., Montzka et al., 2011;Naus et al., 2019;Rigby et al., 2017;Turner et al., 2017).These inversions are often under constrained (Turner et al., 2017), leading to disagreement among studies in OH variability (Patra et al., 2021).Finally, the global decline in MCF could soon necessitate an alternative methodology to infer OH abundance (Liang et al., 2017).
Chemistry transport models can also inform understanding of OH trends and variability, although model intercomparison projects (MIPs) highlight some limitations with this approach.While most models participating in the Chemistry Climate Model Initiative MIP simulated a global increase in tropospheric OH of 0.1-0.3× 10 5 molecules/cm 3 from 2000 to 2010, there was little agreement in the interannual variability (IAV) of global OH and in the spatial distribution of the trends (Zhao et al., 2019).Likewise, global OH generally increased from 2000 to 2015 in models that participated in AerChemMIP (Stevenson et al., 2020).This disagrees, however, with observationally-constrained studies (e.g., Nicely et al., 2018;Rigby et al., 2017;Turner et al., 2017), where OH decreases or remains constant over the same period.
Satellite proxy data are one promising new pathway for constraining OH trends and variability.Polar-orbiting satellites can provide global coverage at relatively high resolution, allowing for the determination of OH trends and variability at significantly finer scales than MCF inversions.Previous work has used Ozone Monitoring Instrument (OMI) formaldehyde (HCHO) to constrain OH over the global oceans (Wolfe et al., 2019), and OMI nitrogen dioxide (NO 2 ) and HCHO to estimate surface OH in North American urban areas (Zhu et al., 2022a).Zhu et al. (2022b) then used this satellite-constrained OH to determine trends in urban OH and their relationship to ozone (O 3 ) production.Pimlott et al. (2022) used multiple retrievals from IASI (Infrared Atmospheric Sounding Interferometer) to constrain OH globally between 600 and 700 hPa and investigated the role of various drivers on IAV on the hemispheric scale.Finally, Anderson et al. (2023) combined observations from multiple satellites to produce a tropospheric column OH (TCOH) product over the tropical oceans.
Our goal is to demonstrate the utility of the Anderson et al. (2023) TCOH product in identifying trends and variability in tropical, tropospheric OH and in highlighting the role of the chemical and dynamical drivers of that variability.We focus our work on boreal autumn (September-November; SON) as this season has large spatial variability in TCOH trends and modeling studies demonstrate strong impacts on TCOH IAV from the El Niño Southern Oscillation (ENSO) over oceans (e.g., Naus et al., 2021;Turner et al., 2018).As we are aiming to highlight the utility of this method, we limit our analysis to the tropics, where ENSO impacts on OH and global CH 4 oxidation maximize (Anderson et al., 2021).To our knowledge, this is the first observationally-based study examining TCOH trends and variability in this region at high resolution (1°× 1°), representing a significant advance.
In Section 2, we describe the TCOH product as well as a multiple linear regression (MLR) model.In Section 3, we examine the tropical TCOH distribution and variability, TCOH trends and their underlying causes (Section 3.1), and TCOH anomalies during the 2015 El Niño (Section 3.2).

Satellite-Constrained TCOH Product and MLR Model Descriptions
In this section, we briefly describe the TCOH product and the underlying satellite data (Section 2.1) as well as a MLR model (Section 2.2) used for understanding trends and the relative impacts of ENSO and the Indian Ocean Dipole (IOD) on TCOH IAV.

Description of the Satellite-Constrained TCOH Product
We use Gradient Boosted Regression Trees (GBRT) (T.Chen & Guestrin, 2016), a machine learning technique, to create the satellite-constrained TCOH product.Currently, the GBRT model is valid only for tropical regions (i.e., latitudes equatorward of 25°) over the ocean.Future work will expand this methodology globally.Input variables to the model include tropospheric column NO 2 , carbon monoxide (CO), HCHO, and water vapor (H 2 O (v) ), total column O 3 , aerosol optical depth (AOD) at 550 nm, solar zenith angle, latitude, sea surface temperature (SST), and H 2 O (v) at 7 layers (Table S1 in Supporting Information S1).We trained the model on output from the MERRA-2 GMI simulation (Modern Era Retrospective analysis for Research and Applications, version 2, Global Modeling Initiative) (Strode et al., 2019) of the NASA GEOS (Goddard Earth Observing System) model (e.g., Pawson et al., 2008).The GBRT model reproduces independent output from MERRA2-GMI within approximately 5% and captures the variability of OH columns derived from in situ observations (Anderson et al., 2023).
To calculate TCOH, we input Level 3 (L3) satellite retrievals (Table S1 in Supporting Information S1) of the necessary variables into the GBRT model.All retrievals are the same as those described in Anderson et al. (2023) except for HCHO.We now use the version 4 OMI Smithsonian Astrophysics Observatory retrieval (Text S1 in Supporting Information S1).We used area weighting to average all satellite data to monthly resolution and a common 1°× 1°grid before inputting into the GBRT model.We then averaged the resultant TCOH for SON.Because the L3 products are cloud filtered, generally omitting cloud fractions greater than 30%, the TCOH product is representative of mostly clear sky conditions.Data span from 2005 to 2019, the period for which retrievals for all GBRT model inputs are available.We omit data from 2009, where observations from some satellites are missing for September.A more complete description of the GBRT model and the TCOH product is in Text S2 in Supporting Information S1 and Anderson et al. (2023).

Description of the MLR Model
To determine trends in TCOH as well as the influence of the ENSO and IOD on TCOH variability, we developed a MLR model following the methodology of Ziemke et al. (1997).The MLR has the form: where α, β, γ, and δ are constant regression coefficients and t is time.Because we examine only SON, we do not include a seasonal time dependence as described in Ziemke et al. (1997).The TCOH trend is given by the coefficient β.For the ENSO and IOD terms, we use the Niño3.4index and the Dipole Mode Index, respectively, both calculated with the SST dataset described in Hirahara et al. (2014).We normalized and detrended both the ENSO and IOD time series prior to analysis.Detrending allows for any trend in the ENSO or IOD terms to be accounted for in the β term.R(t) is the residual.
We used the MLR to determine the trend, as well as the ENSO and IOD coefficients, across the spatial domain.
We applied the model to each grid box individually and only include, in the analysis below, grid boxes with at least 10 years of data.

Observationally-Constrained TCOH Trends and Variability in the Tropics
The climatological mean TCOH distribution (Figure 1a) for SON exhibits significant spatial heterogeneity.The domain-wide mean is 6.1 × 10 12 molecules/cm 2 , with columns ranging from 4.2-9.5 × 10 12 molecules/cm 2 .Maxima are primarily located in the Northern Hemisphere in regions where anthropogenic emissions dominate continental outflow (e.g., coastal India, China, Southeast Asia, and Central America).This proximity to coastlines is consistent with the short OH lifetime as well as the comparatively high abundance of OH drivers over land, particularly NO 2 (e.g., Duncan et al., 2016).Also evident is the impact of NO X shipping emissions, evidenced by TCOH maxima along shipping lanes between Sri Lanka and Malaysia (e.g., Richter et al., 2004).Another broad TCOH maximum extends across the Atlantic Ocean, a likely result of influence from African and South American biomass burning.Local minima are primarily found at the southernmost extent of the study domain, where advection of comparatively clean air from high latitudes has more influence than in the deep tropics.
In contrast to the climatological mean, TCOH variability, defined as 1σ of the distribution of an individual grid box, maximizes around the maritime continent (MC) and regions further from the coastline (Figure 1b).Variability in these regions is generally on the order of 10%-15% of the climatological mean, compared to less than 5% in other areas.Many of the regions of maximum variability (e.g., Indian Ocean, South Pacific Convergence Zone, and Eastern Equatorial Pacific Ocean) correspond with areas with frequent, deep convection (e.g., Pfister et al., 2022).The co-location of the maxima in variability and these convective regions suggests a possible role for dynamical transport in controlling OH variability.Similarly, the variability around Indonesia and over the Indian Ocean implies a possible role for ENSO and the IOD (e.g., Anderson et al., 2021;Naus et al., 2021).We briefly discuss possible cloud impacts on the TCOH product in Section 3.1.
In the following sections, we explore the spatial heterogeneity of the OH distribution and variability by examining the underlying causes of trends in TCOH (Section 3.1) in four regions and the impact of the 2015 El Niño on TCOH distributions (Section 3.2).In this work, we aim to highlight large-scale and interesting features and do not attempt to explain all aspects of TCOH trends and variability, a topic for future study.

Understanding the Magnitude and Spatial Variability of TCOH Trends
TCOH trend distributions exhibit large spatial heterogeneity (Figure 2), highlighting an advantage of this methodology over MCF inversions.Positive trends of up to 15%/decade dominate over much of the domain, particularly over the Eastern Pacific Ocean, although there are large, coherent regions of decreases in TCOH, generally between 2.5% and 5%/decade, particularly over the Northern Hemispheric Atlantic and Indian Oceans.
When averaged domain-wide these distinctions are lost, resulting in a positive trend of 2.1%/decade.
We now separately examine the underlying causes for trends in four broad regions: the South Atlantic Ocean (SAtl), the North Atlantic Ocean (NAtl), the MC, and the Eastern Pacific Ocean (EPac).These areas encompass a range of underlying causes of TCOH trends and highlight various strengths and weaknesses of our methodology.
To infer the relative influence of each satellite retrieval used in the GBRT model, we recalculated TCOH, individually replacing the values of each input for all years with its climatological mean.For H 2 O (v) , we replace the column and layer observations simultaneously.We then take the difference between the actual trend and the trend calculated with the climatological mean to infer the relative contribution of each species.We show the species with the largest impact on the TCOH trends in Figure 3 and those with limited impact in Figure S1 of Supporting Information S1.We note that HCHO is not necessarily a driver of OH but is itself a proxy for the impacts of volatile organic compounds (VOCs) on OH.Over the oceans, HCHO is generally strongly correlated with OH (Baublitz et al., 2023;Wolfe et al., 2019), suggesting that decreasing OH trends should be accompanied by decreases in HCHO.
The increasing TCOH trend over the SAtl, which is more than 7.5%/decade in some regions, is likely being driven by changes in South American emissions.Decreasing CO in the SAtl (Figure S2f in Supporting Information S1) accounts for approximately half of the observable TCOH trend in the region.Because CO is the dominant tropospheric OH sink (e.g., Lelieveld et al., 2016), trends in CO and TCOH are inversely related.Decreases in MOPITT (Measurement of Pollution in the Troposphere) CO, Quick Fire Emissions Dataset CO emissions, OMI NO 2 , and MODIS (Moderate Resolution Imaging Spectrometer) AOD (Figure S2 in Supporting Information S1) over the Brazilian Amazon during this period suggest that a decline in South American biomass burning is the underlying cause of these CO changes, in agreement with Buchholz et al. (2021) and Y. Chen et al. (2023).The HCHO trend downwind of the biomass burning region (Figure 3b and Figure S2b in Supporting Information S1) is consistent with this idea as a decrease in fires would lead to decreases in emissions of HCHO and its precursors (e.g., De Smedt et al., 2015).This decrease in HCHO and likely its VOC precursors partially offset the increasing TCOH trend in the region (Figure 3b).
Increases in NO 2 and H 2 O (v) also lead to increases in TCOH over the SAtl.While NO 2 columns in Brazilian biomass burning regions are decreasing, there is a comparatively large increasing trend in NO 2 over southeastern Brazil (Figure 3a), where biomass burning is more limited, indicating increases in other anthropogenic NO X sources.These changes likely contribute to increases in TCOH over the SAtl downwind of this region.Finally, Changes in VOCs are likely related to the decreasing trend over the NAtl.Formaldehyde, a VOC proxy, is the only GBRT input associated with decreases in TCOH in the region.While HCHO columns decrease over the NAtl as well as in upwind regions (Figure 3b and S2b in Supporting Information S1), the cause of these trends is unknown.In the remote atmosphere, CH 4 is the dominant HCHO source (e.g., Anderson et al., 2017).Methane has increased globally over the study period (Turner et al., 2019), however, suggesting an offsetting Because of its short lifetime, long-range transport of HCHO is unlikely (Vigouroux et al., 2009), although a reduction of emissions in HCHO precursors, such as longer-lived VOCs, could explain the trend, and is consistent with the trends in HCHO over the surrounding land regions.
The trends around the MC and surrounding regions of the Pacific and Indian Oceans likely result from changes in biomass burning emissions but are also potentially influenced by changes in other Indonesian emissions.Indonesian NO 2 abundance and biomass burning emissions have generally decreased over 2005-2019, leading to a decreasing trend in TCOH immediately around Indonesia (Figure 3a).Lofting and subsequent transport of these emissions through Walker Circulations to portions of the Indian and Pacific Oceans likewise contribute to the decreasing TCOH trend in these regions.In contrast, HCHO increases over the MC.Since this is anticorrelated with the NO 2 and biomass burning trend, it is possible this is related to increases in other anthropogenic or biogenic VOC emissions.The relationship between HCHO and the TCOH trend over the Indian and Pacific Oceans seems decoupled from the concentration trends over the MC, as HCHO and TCOH both decrease in these regions.H 2 O (v) leads to a decrease in TCOH around the MC, primarily driven by decreasing H 2 O (v) trends in the middle troposphere (Figure S2e in Supporting Information S1).This is consistent with B. Chen and Liu (2016), who found an anti-correlation between surface temperature and tropospheric water vapor in the region which they attribute to a large moisture flux divergence.The largest trends in TCOH, above 7.5%/decade in some regions, are off the coast of Central America in the EPac, where NO 2 is the dominant contributor.While NO 2 in this region does increase, with the exception of one local maximum that does not correspond to the maximum in the TCOH trend, the magnitude of that trend is comparable to much of the rest of the domain (Figure S2a in Supporting Information S1), where the impact of NO 2 on the TCOH trend is lower.A unique photochemical environment could account for a different sensitivity of OH to NO 2 in this region.Alternatively, retrieval errors could be an issue, as this region coincides with an area of deep convection (Pfister et al., 2022) and frequent clouds could affect the retrievals, although other regions of deep convection in this study do not exhibit similar features.Further study is needed to understand trends in this region.
This methodology does have several minor limitations.Because we omit CH 4 from our GBRT model, we cannot discern the impacts of recent increases in CH 4 (Turner et al., 2019) on TCOH.Given the long CH 4 lifetime, however, any impacts on TCOH trends would likely have limited spatial variability.Further, because GBRT models are not process-based and OH chemistry is highly non-linear, it is also difficult to state exactly the quantitative contribution of an individual input to OH changes.Therefore, we primarily limit discussion to qualitative impacts of the changes of the various GBRT model inputs.Finally, because we use L3 products for this analysis, differences in averaging among the retrievals from the pixel to the gridded level could affect results.For example, in cloudy regions with missing data, some L3 products could use a single pixel to represent a value over one much larger grid box.This could lead to artifacts such as unrealistically large variability.Because of differences in data products, it is not possible to quantify these effects for all retrievals.Future work will incorporate Level 2 retrievals to limit these differences.

Using the 2015 El Niño to Inform Understanding of Drivers of OH Variability
We now investigate the impacts of El Niño on TCOH interannual variability with a case study of the 2015 event.MLR analysis (Figure 2b) demonstrates an anti-correlation between the Niño3.4Index and TCOH over much of the domain, although there are several regions with coherent positive correlation, suggesting that tropical TCOH should, on average, decrease during an El Niño.This relationship is somewhat confounded by an opposing response to the IOD (Figure 2c), which is itself correlated (r = 0.43) with ENSO.The global decrease is in agreement with previous observationally constrained global (Patra et al., 2021) and regionally averaged (Pimlott et al., 2022) studies.Indeed, the largest domain-wide minima in TCOH occurred during two El Niños (2006 and2015), with the detrended anomaly during 2015 a factor of two larger than any other year (Figure S3 in Supporting Information S1).We focus our analysis on this exceptional event.
To determine the inferred contribution for each input to the spatial variability (Figure 4a) in the TCOH anomalies, we use the GBRT runs described in Section 3.1, where we hold individual inputs constant to their climatological mean.We then detrend the data and calculate the difference in the 2015 anomaly between the baseline and climatological mean GBRT runs (Figure 4).El Niño-induced changes in convection and subsequently H 2 O (v) lead to a dipole pattern of TCOH anomalies over parts of the domain.Changes in the Walker Circulation over the MC lead to reductions in convection, H 2 O (v) (Figure S4c in Supporting Information S1), and decreased primary production of OH (Figure 4b), while the effects are reversed in the central Pacific.This dipole pattern is consistent with the ENSO effect on O 3 (Oman et al., 2013).The OH response to O 3 is relatively small in the region, however.This is likely a result of the GBRT model using total column O 3 as an input and the relatively low perturbation to the total column caused by tropospheric changes.
In contrast, Indonesian biomass burning emissions lead to domain-wide decreases in OH through a strong positive CO anomaly.Increased fires over Sumatra and Kalamantan during the El Niño resulted in the largest CO anomalies (Figure S4a in Supporting Information S1) of the satellite record (Field et al., 2016).Given the CO lifetime and its role as the dominant OH sink, this increase in fires leads to OH decreases across the tropics.NO X and VOC emissions from these fires have a limited impact on OH anomalies outside of areas immediately surrounding Indonesia, however, which is reasonable given their shorter lifetimes.
In the equatorial Indian and Atlantic, large negative anomalies result from a combination of effects.Over the Atlantic, negative anomalies in H 2 O (v) , HCHO, and NO 2 align with the increased CO from Indonesian fires to produce this anomaly (Figure 4 and Figure S4 in Supporting Information S1).Over the Indian Ocean, positive anomalies in H 2 O (v) have the opposite effect, as the Walker Circulation during the positive phase of the IOD increases convection over the western portion of the basin.The cause for the strong HCHO anomaly in both basins is unknown.The increased drought and biomass burning associated with an El Niño should lead to increases in emissions of HCHO precursors (e.g., Morfopoulos et al., 2022), although extreme temperatures and drought could have the opposite effect.Further study is needed to understand these anomalies.

Summary and Conclusions
We have demonstrated the ability of a satellite-constrained OH product to capture spatially resolved trends and variability.While tropical OH is increasing at a rate of 2.1%/decade domain-wide, there is substantial spatial variability, with large regions in the Eastern Pacific increasing by 7.5%/decade or higher and areas in the North Atlantic decreasing at rates of up to 5%/decade.Similarly, while previous observationally-constrained studies have shown negative anomalies in OH during the 2015 El Niño, we also demonstrate that these anomalies are not universal, with a prominent dipole pattern over the MC and central Pacific, for example, We also demonstrate the ability of the GBRT methodology to help qualitatively attribute the relative impacts of the model inputs to OH trends and variability.
This ability to produce spatially resolved trends and variability can act as a complement to existing methods of constraining OH, such as MCF inversions.Expanding the analysis performed here to other seasons will provide additional context to the annual resolution of the MCF inversions.Future plans to expand the methodology to the extratropics and over land will allow for comparisons to MCF-derived OH trends, while plans to include satellite retrievals of additional species, such as peroxyacetyl nitrate, to the GBRT model could help improve understanding of the behavior of OH in chemical models (e.g., Murray et al., 2021).Given the lack of directly observed OH and the uncertainties associated with interpreting CH 4 trends and IAV, the satellite product could also provide additional needed constraints on the CH 4 budget.Finally, we are unable to fully validate the results found here because of lack of direct OH observations.Additional field campaigns measuring tropospheric columns of OH and its drivers are needed to further constrain the variability of OH as well as its drivers and proxies.

Figure 1 .
Figure 1.The mean (a) and standard deviation (b) of tropospheric column OH for SON is shown for 2005-2019.

Figure 2 .
Figure 2. (a) The trend in tropospheric column OH (TCOH) (β), (b) El Niño Southern Oscillation coefficient (γ), and (c) Indian Ocean Dipole coefficient (δ) from the multiple linear regression.The contours in panel (a) indicate the trend as a percent of the climatological mean TCOH over 2005-2019 in %/decade.Only statistically significant trends, defined as grid boxes where the trend is greater than the 2σ uncertainty in β, are shown.

Figure 3 .
Figure 3.Over the oceans, we show the inferred contribution of (a) NO 2 , (b) HCHO, (c) CO, and (d) H 2 O (v) to the total OH trend over 2005-2019.Only regions where the inferred contribution to the trend is greater than 25% is shown.Over land, we show the trends in (a) Ozone Monitoring Instrument (OMI) NO 2 , (b) OMI HCHO, and (c) MOPITT CO for the same time period.We use a simple linear regression to determine trends over land and report them as a percentage of the mean climatological value.

Figure 4 .
Figure 4.The anomaly in detrended tropospheric column OH for (a) 2015.The inferred contribution of (b) H 2 O, (c) CO, (d) total column O 3 , (e) NO 2 , and (f) HCHO to the anomaly shown in panel (a).