Proceedings Article | 19 December 2008
KEYWORDS: Optimization (mathematics), Remote sensing, Absorption, Algorithm development, Ocean optics, Reflectivity, Statistical analysis, Evolutionary algorithms, Particles, Backscatter
During recent years, more and more efforts have been focused on developing new models based on ocean optics theory
to retrieve water's bio-geo-chemical parameters or inherent optical properties (IOPs) from either ocean color imagery or
in situ measurements. Basically, these models are sophisticated, and hard to invert directly, look up table (LUT)
technique or optimization methods are employed to retrieve the unknown parameters, e.g., chlorophyll concentration,
CDOM absorption, etc. Many researches prefer to use time-consuming global optimization methods, e.g., genetic or
evolutionary algorithm, etc. In this study, different optimization methods, smooth nonlinear optimization (NLP), global
optimization (GO), nonsmooth optimization (NSP), are compared based on the sophisticated hyper-spectral semianalytical
(SA) algorithm developed by Lee et al., retrieval accuracy and performance are evaluated. It is found that
retrieval accuracy don't have much difference, the performance difference, however, is much larger, NLP works very
well for the SA model. For a given model, it is better to analyze the model is linear, nonlinear or nonsmooth category
problem, sometimes, convex also need to be determined, or linearize some nonsmooth problem caused by if decision,
then select the corresponding category optimization methods. Initial values selection is a big issue for optimization, the
simple statistical models (e.g., OC2 or OC4) are used to retrieve the unknowns as initial values.
