Soft Computing-based Ocean Color Remote Sensing

Like other inverse problems, the solutions to inverse of ocean color fall into two categories: implicit and explicit. Explicit solutions are inverse transfer functions that give the colored water constituents as functions of the upwelling light spectrum from the water body. These inverse transfer functions were obtained traditionally by empirical, semi-analytical and analytical approaches. The semi-analytical and analytical approaches are computationally intensive and generally produced lower accuracy, and the uncertain and nonlinear relationships between radiance and constituent concentrations make it difficult to construct empirical transfer functions to accurately retrieve constituent concentrations. Implicit solutions use forward optical models to simulate the spectrum signal and optimization procedures to minimum the deviation between the measured and computed spectrum, but almost previous applications of this approach suffer from their local search techniques. The aim of this dissertation is to apply some data analysis techniques of soft computing, which is an emerging approach to exploit the tolerance for imprecision, uncertainty and partial truth to achieve tractability, robustness and better rapport with reality. These techniques include two modeling approaches and one optimization approach. The first modeling approach is artificial neural network(NN). Three layer feed-forward NNs for the inversion of chlorophyll concentrations from remote sensing reflectance measurements were designed and trained on a subset of the SeaBAM data set, using the Levenberg-Marquardt training algorithm which in combination with Bayesian regularization. The remaining SeaBAM data set was then applied to evaluate the performances of NNs and compared with those of the SeaBAM empirical algorithms. NNs achieved better inversion accuracy than the empirical algorithms in most of chlorophyll concentration range, especially in the intermediate and high chlorophyll regions and Case 2 waters. Systematic overestimation existed in the very low chlorophyll (<0.031mg/m3) region, and little improvement was obtained by changing the size of the training data set. The inversion results showed that NN is a feasible and universal method for inversion of chlorophyll concentration. Adding to its ability to model nonlinear transfer function, the most important advantage of NN is that it can use information of all bands directly instead of in frozen and a priori way as the empirical methods do, which makes it can perform well under various circumstances such as in the Caseâ…¡waters. However, NNs are not perfect for modeling transfer function practically. They...