Technical note: Unsupervised classification of ozone profiles in UKESM1

Abstract

Abstract. The vertical distribution of ozone in the atmosphere, which features complex spatial and temporal variability set by a balance of production, loss, and advection, is relevant for both surface air pollution and climate via its role in radiative forcing. At present, the way in which regions of coherent ozone structure are defined relies on somewhat arbitrarily drawn boundaries. Here we consider a more general, data-driven method for defining coherent regimes of ozone structure. We apply an unsupervised classification technique called Gaussian mixture modeling (GMM), which represents the underlying distribution of ozone profiles as a linear combination of multi-dimensional Gaussian functions. In doing so, GMM identifies coherent groups or subpopulations of the ozone profile distribution. As a proof-of-concept study, we apply GMM to ozone profiles from three subsets of the UKESM1 coupled climate model runs carried out for CMIP6: specifically, the seasonal mean of a historical subset (2009–2014) and two subsets from two different future climate projections (i.e., SSP1-2.6 and SSP5-8.5). Despite not being given any spatiotemporal information, GMM identifies several spatially coherent regions of ozone structure. Using a combination of statistical guidance and post hoc judgment, we select a six-class representation of global ozone, consisting of two tropical classes and four mid-to-high-latitude classes. The tropical classes feature a relatively high-altitude tropopause, while the higher-latitude classes feature a lower-altitude tropopause and low values of tropospheric ozone, as expected based on broad patterns observed in the atmosphere. Both of the future projections feature lower ozone concentrations at 850 hPa than the historical benchmark, with signatures of ozone hole recovery. We find that the area occupied by the tropical classes is expanded in both future projections, which are most prominent during austral summer. Our results suggest that GMM may be a useful method for identifying coherent ozone regimes, particularly in the context of model analysis.

Ozone Profile Retrieval Algorithm Based on GEOS-Chem Model in the Middle and Upper Atmosphere

“…The concentration of ozone is controlled by a series of chemical, radiative, and kinetic processes across a range of spatial and temporal scales. That means that ozone is generated by gas-phase photochemical reactions and is destroyed by reactions with chlorine, nitrogen, hydrogen, and bromine radicals [55]. The source and sink concentrations were also inconsistent across different regions and altitudes, while photochemical reaction rates depend on specific temperatures.…”

A novel heuristic method for detecting overfit in unsupervised classification of climate model data

“…Unsupervised classification methods, that is, methods that do not know a priori what the properties of these groups might be, have proven adept at identifying coherent spatial structures within climate data, even when no spatial information is supplied to the algorithm. In studies of ocean and atmospheric data, two commonly used unsupervised classification methods are k-means (Solidoro et al, 2007;Hjelmervik and Hjelmervik, 2013;Hjelmervik et al, 2015;Sonnewald et al, 2019;Houghton and Wilson, 2020;Yuchechen et al, 2020;Liu et al, 2021) and Gaussian mixture modeling (GMM) (Hannachi and O'Neill, 2001;Hannachi, 2007;Tandeo et al, 2014;Maze et al, 2017a;Jones et al, 2019;Crawford, 2020;Sugiura, 2021;Zhao et al, 2021;Fahrin et al, 2022). K-means attempts to find coherent groups by "cutting" the abstract feature space using hyperplanes, whereas GMM attempts to represent the underlying covariance structure in abstract feature space using a linear combination of multi-dimensional Gaussian functions.…”

“…The most commonly used statistical criterion is the minimum in the Bayesian information criterion (BIC; Schwarz, 1978), used in Fahrin et al (2022), Sugiura (2021), Zhao et al (2021), Jones et al (2019), Maze et al (2017b), Hjelmervik (2013, 2014), Hjelmervik et al (2015), and Sonnewald et al (2019). The BIC is comprised of two terms: a term that rewards the statistical likelihood of the model and a term that penalizes overfitting.…”