A novel ozone profile shape retrieval using full-physics inverse learning machine (FP-ILM)

Abstract

Identifying ozone profile shapes from nadir-viewing satellite sensors is a critical yet challenging task for accurate reconstruction of vertical distributions of ozone relevant to climate change and air quality. Motivated by the need to develop a methodology to fast, reliably, and efficiently exploit ozone distributions and inspired by the success of machine learning, this paper introduces a novel algorithm for estimating ozone profile shapes from satellite ultraviolet absorption spectra. The Full-Physics Inverse Learning Machine (FP-ILM) algorithm successfully characterizes ozone profile shapes using machine learning approaches. Its implementation mainly consists of a clustering process based on a semi-supervised agglomerative algorithm, a classification process based on full-physics radiative transfer simulations and a neural network (NN), and a profile scaling process based on a NN ensemble. The classification model has been trained with synthetic data generated by a forward model in conjunction with “smart sampling,” while the scaling model corresponding to each cluster requires total ozone information. The main innovation of FP-ILM is that, unlike conventional inversion methods, the ozone profile retrieval is formulated as a classification problem, leading to a noteworthy speed-up and accuracy when dealing with applications of satellite data. An outstanding retrieval performance with errors of less than 10% over 100–1 hPa has been obtained for synthetic measurements. Furthermore, the ozone profiles retrieved from the Global Ozone Monitoring Experiment–2 data using FP-ILM and the optimal estimation method reach an encouraging agreement (the differences are less than 6 Dobson Units or within 5%–20%).