An ensemble neural network atmospheric correction for Sentinel-3 OLCI over coastal waters providing inherent model uncertainty estimation and sensor noise propagation

Abstract

Accurate atmospheric correction (AC) is a prerequisite for quantitative ocean colour remote sensing and remains a challenge in particular over coastal waters. Commonly AC algorithms are validated by establishing a mean retrieval error from match-up analysis, which compares the satellite-derived surface reflectance with concurrent ground radiometric observations. Pixel-based reflectance uncertainties however, are rarely provided by AC algorithms and those for the operational Ocean and Land Colour Instrument (OLCI) marine reflectance product are not yet recommended for use. AC retrieval errors and uncertainties directly determine the quality with which ocean colour products can be estimated from the marine surface reflectance. Increasingly there is also the need for reflectance uncertainty products to be used as data assimilation inputs into biogeochemical models.

This paper describes the development of a new coastal AC algorithm for Sentinel-3 OLCI that provides pixel-based estimation of the inherent model inversion uncertainty and sensor noise propagation. The algorithm is a full-spectral model-based inversion of radiative transfer (RT) simulations in a coupled atmosphere–ocean system using an ensemble of artificial neural networks (ANN) that were initialized differently during the training process, but composed of the same network architecture. The algorithm has been validated against in-situ radiometric observations across a wide range of optical water types, and has been compared with the latest EUMETSAT operational Level 2 processor IPF-OL-2 v7.01.

In this analysis we found that the ensemble ANN showed improved performance over the operational Level 2 processor with a band-averaged (412–708 nm) mean absolute percentage error (MAPE) of 16% compared to 37% and a four-times lower band-averaged bias of -0.00045 sr-1.

In the ensemble inversion process we account for three uncertainty components: (1) the total model variance that describes the variance of the data from the different ANNs, (2) the prediction variance of the mean, which is based on calculations of the RT simulations and (3) the instrument noise variance of the mean by propagating the OLCI spectral signal-to-noise ratios (SNR). To study algorithm performance and to quantify the contribution of the different uncertainty components to the total uncertainty, we applied the algorithm to an optically complex full resolution (FR) test scene covering coastal waters of the Great Barrier Reef, Australia. The uncertainties associated with the instrument noise variance were found to be two orders of magnitude lower than the uncertainty components of the prediction and total model variances. The overall largest uncertainty component in our uncertainty framework is attributed to the total model inversion error from averaging the responses of the slightly different adapted networks in the ensemble. The algorithm is made publicly available as a Python/C plugin for the Sentinel Application Platform (SNAP).