MEETC2: Ocean color atmospheric corrections in coastal complex waters using a Bayesian latent class model and potential for the incoming sentinel 3 — OLCI mission

Highlights

• Innovative atmospheric correction of ocean color satellite data in coastal waters

• Bayesian inference using priors on reference spectra of aerosol and water reflectance

• Convergence toward realistic non-negative solution conversely to standard products

• Significant improvements of the estimates compared with the standard processing

Abstract

From top-of-atmosphere (TOA) observations, atmospheric correction for ocean color inversion aims at distinguishing atmosphere and water contributions. From a methodological point of view, our approach relies on a Bayesian inference using Gaussian Mixture Model prior distributions on reference spectra of aerosol and water reflectance. A reference spectrum for the aerosol characterizes the specific signature of the aerosols on the observed aerosol reflectance. A reference spectrum for the water characterizes the specific signature of chlorophyll-a, suspended particulate matters and colored dissolved organic matters on the observed sea surface reflectance. In our Bayesian inversion scheme, prior distributions of the marine and aerosol variables are set conditionally to the observed values of covariates, typically acquisition geometry acquisition conditions and pre-estimates of the aerosol and water reflectance in the near-infrared part of the spectrum. The numerical inversion exploits a gradient-based optimization from quasi-randomized initializations.

We evaluate our estimates of the sea surface reflectance from the MERIS TOA observations. Using the MERMAID radiometric in-situ dataset, we obtain significant improvements in the estimation of the sea surface reflectance, especially for the 412, 442, 490 and 510 nm bands, compared with the standard ESA MEGS algorithm and the a state-of-the-art neural network approach (C2R). The mean gain value on the relative error for the 13 bands between 412 and 885 nm is of 57% compared with MEGS algorithm and 10% compared with the C2R. The water leaving reflectances are used in Ocean Color for the estimation of the chl-a concentration, the colored dissolved organic matters absorption and the suspended particulate matters concentration underlying the potential of such approach to improve the standard level 2 products in coastal areas. We further discuss the potential of MEETC2 for the incoming OLCI/Sentinel 3 mission that will be launched in December 2015.

Introduction

The inversion of Ocean Color signal in coastal areas from top-of-atmosphere (TOA) measurements remains a scientific challenge. This is a crucial point for the ocean color community as many governmental policies such as the European Water Framework directive (WFD) rely on estimation of coastal water quality, itself possibly derived from space-based ocean color measurements (http://ec.europa.eu/environment/water/water-framework/index_en.html). Hence, ocean color inversion is certainly among the highest priority research topics for ocean color community. Different aspects may explain the difficulties encountered in this inversion process. Firstly, the contribution of suspended matters to the reflectance in the near infrared range (700–900 nm) is an issue as many algorithms expect these reflectances to be null. This assumption is called the black pixel hypothesis and relies on the strong natural absorption of the water in this domain (Antoine et al., 2006, be Gordon & Wang, 1994). Secondly, bio-optical modeling, i.e. the estimation of the water-leaving reflectance from the Inherent Optical Properties (IOPs, namely the absorption and backscattering of the sea water constituents) in complex coastal waters is also challenging. Despite accurate physical models exist for open clear waters that cover 85% of the oceans, their derivation for coastal waters is more complex (IOCCG, 2000, Maritorena et al., 2002). Lastly, aerosol and water reflectance spectra may show important correlation in the near infrared, a spectral domain typically used by the standard algorithms to distinguish the two contributions.

As a consequence, available operational standard level-2 reflectance products may perform poorly in coastal areas, and consequently these products are often flagged as anomalous values for such areas (MERIS DPM, 2005). The result for end users is typically that very few observations are available in coastal areas if the standard flags are applied. For available pixels in coastal turbid waters, reflectances in the blue and green bands are often underestimated and may involve physically-meaningless negative values (Goyens et al., 2013, Jamet et al., 2011). Park, De Cauwer, Nechad, and Ruddick (2004) show this strongly affects the relevance of level-2 products for the end users, which typically use water-reflectance spectra as inputs to estimate the chlorophyll-a and the suspended particulate matter concentrations (SPM, Doxaran, Froidefond, Lavender, & Castaing, 2002), or the vertical light attenuation (Jamet et al., 2012, Morel et al., 2006, Saulquin et al., 2013, Wang et al., 2009).

Over the last fifteen years, many coastal water algorithms have been developed to address user's needs for reliable water-reflectance data in coastal areas. Among them, the Schiller and Doerffer (1999) MERIS Case2-Regional (C2R) is based on a non-linear machine learning model, namely a Neural Networks (NN, Krasnopolsky and Schiller, 2003, Schiller and Doerffer, 1999), and estimates water reflectance over turbid areas. The learning paradigm relies on the calibration of a non-linear model to relate the available satellite-derived observations to the geophysical quantity of interest from a training dataset. This training dataset typically consists of a collection of in-situ measurements along with the satellite-derived measurements. This learning-based strategy may suffer from two major drawbacks: weak geophysical and biological interpretability of this ‘black-box’ model and the assumption on the representativity of the training dataset. They may restrict the applicability of the model to a specific region and question its validity with respect to the generally unknown variability of the atmospheric and water conditions.

Here, we develop a Bayesian latent class approach to address these limitations. To our knowledge, Bayesian model mixtures have been seldom explored for ocean color inversion (Frouin & Pelletier, 2014). The key feature of our model is the inversion of water and atmospheric signals from TOA observations using a multi-hypothesis setting. Rather than considering a single model, linear or not, we develop a Bayesian framework where the priors stated as mixture of models. Mixture models are trained both for water and aerosol contributions and lead to the identification of the reference spectrum families characterized by their mean spectrum and the associated covariance matrices. This training phase exploits in-situ data or radiative transfer simulations in the atmosphere and the water (Barker et al., 2008, Berk et al., 1999, Deuzé et al., 1989). Contrary to the machine learning approaches (NN, or Support Vector Regressions, SVR; Burges, 1998), the identified a priori distributions of the water and aerosol variables are directly linked to interpretable water types or atmospheric spectra.

Our inversion scheme, referred to hereafter as MEETC2, is applied to the estimation of water reflectances from the MEdium Resolution Imaging Spectrometer (MERIS) TOA observations. Nevertheless the methodology is generic and may be directly applied to other sensors such as the incoming OLCI sensor embedded onto the sentinel 3 platform. Model calibration and validation involve the MEris MAtchup In-situ Database (MERMAID) radiometric in-situ dataset (Barker et al., 2008). Quantitative comparisons with the standard MEGS v8 (Antoine et al., 2006) and the MERIS C2R Neural Network outputs clearly demonstrate the relevance of our approach.