Title

Neural Network Solutions to the Ocean Optics Inverse Problem

Date of Award

1999

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computing

First Advisor

Adel Ali

Advisor Department

Computing

Abstract

An artificial neural network (ANN) hybrid conjugate gradient/simulated-annealing algorithm is developed and applied to the ocean optics inverse problem. The new algorithm seeks to enhance the capabilities and strengths of deterministic and stochastic learning algorithms while minimizing their individual weaknesses. The hybrid ANN is then used to approximate the inverse relationship between the inherent optical properties (IOPs) of absorption and scattering in a natural ocean water column and the water leaving radiance values, which are influenced by the IOPs. Two different ANN inversion relations, explicit and implicit, approximate the inverse relationship. The explicit method refers to a direct inversion relationship where water leaving radiance values are used as inputs to the ANN and IOPs are used as outputs. An implicit approach is two step: (1) it first uses an ANN to approximate the forward relationship between IOPs and water leaving radiance, (2) once the forward approximation has converged to the desired error tolerance the ANN weights are held constant while the input values are iteratively adjusted to agree with a desired output (water leaving radiance). The two different approaches are used to investigate the degree to which the many-to-one problem affects this particular inverse problem. Several statistical analytic tests are used to determine the degree to which the training data components are correlated with each other. Principal component analysis is applied to the water leaving radiance in an effort to discern the effect multispectral data plays in creating unique IOPs to water leaving radiance mappings. Validation experiments from June 1998 are used to determine the ANN produced inversion approximation effectiveness. An error analysis technique for artificial neural networks is formulated and applied to the trained explicit network. The different error contributions are also considered and discussed.