Statistical Inference in Ocean-Color Remote Sensing

Robert Frouin, Scripps Institution of Oceanography, UCSD, rfrouin@ucsd.edu (Presenter)
Bruno Pelletier, Department of Mathematics, University of Rennes II, bpelletier@ucsd.edu

The ocean color remote sensing problem is studied in the context of Bayesian inverse problems theory, where the solution is formulated as a conditional probability distribution. This probability distribution, called the posterior distribution, describes the likelihood of encountering possible values of the geophysical parameters given the satellite observations. Solving the inverse problem for a given satellite data therefore amounts to constructing this posterior distribution. The task is difficult: a distribution is an infinite-dimensional parameter, and it is common to estimate relevant characteristics of the posterior distributions. In this work, we construct estimates of the conditional expectations and covariance matrices. The retrieved geophysical parameters include water reflectance, atmospheric reflectance, and aerosol optical thickness. For each of these parameters, the mean and covariance of the posterior distribution is approximated using a partition regression estimate, where the partition is achieved via a binary tree. The construction is performed for multiple geometries, resulting in individual inverse models that are combined in a unique global model, valid for an arbitrary geometry, by kernel smoothing. This type of model is in the form of a function field. The covariance estimates allow one to attach an uncertainty to the retrieved marine reflectance. A quality index is also defined to detect retrievals that cannot be trusted, which occurs when the satellite observation differs too much from the forward model. The methodology, when tested on a SeaWiFS match-up dataset, gives estimates of water reflectance in agreement with in situ measurements. Application to SeaWiFS imagery reveals a substantial noise reduction in the spatial fields of water reflectance compared with the SeaDAS-derived fields.

Presentation Type:  Poster

Session:  Other   (Wed 10:00 AM)

Associated Project(s): 

• Frouin, Robert: Fields of Nonlinear Regression Models for Inversion of Satellite Ocean-Color Data ...details

Poster Location ID: 158