A new method (M&amp;lt;sup&amp;gt;3&amp;lt;/sup&amp;gt;Fusion v1) for combining observations and multiple model output for an improved estimate of the global surface ozone distribution

Kai-Lan Chang,Béatrice Josse,Meiyun Lin,Makoto Deushi,Martin G Schultz,Christoph A Keller,Virginie Marécal,Marc L Serre,Junhua Liu,Owen R Cooper,J Jason West,Kengo Sudo

doi:10.5194/gmd-12-955-2019

Abstract

Abstract. We have developed a new statistical approach (M3Fusion) for combining surface ozone observations from thousands of monitoring sites around the world with the output from multiple atmospheric chemistry models to produce a global surface ozone distribution with greater accuracy than can be provided by any individual model. The ozone observations from 4766 monitoring sites were provided by the Tropospheric Ozone Assessment Report (TOAR) surface ozone database, which contains the world's largest collection of surface ozone metrics. Output from six models was provided by the participants of the Chemistry-Climate Model Initiative (CCMI) and NASA's Global Modeling and Assimilation Office (GMAO). We analyze the 6-month maximum of the maximum daily 8 h average ozone value (DMA8) for relevance to ozone health impacts. We interpolate the irregularly spaced observations onto a fine-resolution grid by using integrated nested Laplace approximations and compare the ozone field to each model in each world region. This method allows us to produce a global surface ozone field based on TOAR observations, which we then use to select the combination of global models with the greatest skill in each of eight world regions; models with greater skill in a particular region are given higher weight. This blended model product is bias corrected within 2∘ of observation locations to produce the final fused surface ozone product. We show that our fused product has an improved mean squared error compared to the simple multi-model ensemble mean, which is biased high in most regions of the world.

Highlights

Tropospheric ozone is a pollutant detrimental to human health and has been associated with a range of adverse cardiovascular and respiratory health effects due to short-term and long-term exposure (World Health Organization, 2005; Jerrett et al, 2009; US Environmental Protection Agency, 2013; GBD, 2015; Turner et al, 2016; Cohen et al, 2017)
We focus on the annual maximum of the 6-month running mean of the maximum daily 8 h average (DMA8) at every site in the Tropospheric Ozone Assessment Report (TOAR) database
We carry out the spatial interpolation by using the combination of the integrated nested Laplacian approximation (INLA) framework (Rue et al, 2009) and the stochastic partial differential equation (SPDE) technique (Lindgren et al, 2011), available as an R package (Lindgren and Rue, 2015). The details of this technique are rather complex and the reader is referred to the original paper (Lindgren et al, 2011); we describe the key component of this INLA-SPDE technique in Appendix A

Summary

Introduction

Tropospheric ozone is a pollutant detrimental to human health and has been associated with a range of adverse cardiovascular and respiratory health effects due to short-term and long-term exposure (World Health Organization, 2005; Jerrett et al, 2009; US Environmental Protection Agency, 2013; GBD, 2015; Turner et al, 2016; Cohen et al, 2017). A useful endeavor for producing an accurate representation of the global surface ozone distribution is to combine the output from many models in a way that takes advantage of the strengths of each model and minimizes the weaknesses Such efforts have already been made for both climate and chemistry–climate models. Multi-model output has been combined using a parametric approach, either by assigning an equal or optimum weight to each model (Stevenson et al, 2006; He and Xiu, 2016; Braverman et al, 2017) or by tuning the initial conditions under different scenarios or parameterizations (Cariolle and Teyssèdre, 2007; Wu et al, 2008; Young et al, 2013) These approaches often assume that individual model biases will at least partly cancel by averaging or weighting, according to certain measures of predictive performance. The combined model product is likely to be more accurate than a single model prediction, as has been shown for multi-model combinations of past or present-day climate (Buser et al, 2009; Knutti et al, 2010; Weigel et al, 2010; Chandler, 2013)

Objectives

Results

Conclusion