Bayesian methodology for inverting satellite ocean-color data

The inverse ocean color problem, i.e., the retrieval of marine reflectance from top-of-atmosphere (TOA) reflectance, is examined in a Bayesian context. The solution is expressed as a probability distribution that measures the likelihood of encountering specific values of the marine reflectance given the observed TOA reflectance. This conditional distribution, the posterior distribution, allows the construction of reliable multi-dimensional confidence domains of the retrieved marine reflectance. The expectation and covariance of the posterior distribution are computed, which gives for each pixel an estimate of the marine reflectance and a measure of its uncertainty. Situations for which forward model and observation are incompatible are also identified. Prior distributions of the forward model parameters that are suitable for use at the global scale, as well as a noise model, are determined. Partition-based models are defined and implemented for SeaWiFS, to approximate numerically the expectation and covariance. The ill-posed nature of the inverse problem is illustrated, indicating that a large set of ocean and atmospheric states, or pre-images, may correspond to very close values of the satellite signal. Theoretical performance is good globally, i.e., on average over all the geometric and geophysical situations considered, with negligible biases and standard deviation decreasing from 0.004 at 412 nm to 0.001 at 670 nm. Errors are smaller for geometries that avoid Sun glint and minimize air mass and aerosol influence, and for small aerosol optical thickness and maritime aerosols. The estimated uncertainty is consistent with the inversion error. The theoretical concepts and inverse models are applied to actual SeaWiFS imagery, and comparisons are made with estimates from the SeaDAS standard atmospheric correction algorithm and in situ measurements. The Bayesian and SeaDAS marine reflectance fields exhibit resemblance in patterns of variability, but the Bayesian imagery is less noisy and characterized by different spatial de-correlation scales. Experimental errors obtained from match-up data are similar to the theoretical errors determined from simulated data. Regionalization of the inverse models is a natural development to improve retrieval accuracy, for example by including explicit knowledge of the space and time variability of atmospheric variables. (C) 2014 Elsevier Inc. All rights reserved.