A hybrid neural network algorithm for hyperspectral, remotely sensed, shallow-water bathymetry

David Clarence Hughes


The purpose of this research is to test a novel neural network methodology capable of managing multiple ocean floor bottom types and water column depths to solve the ocean inverse problem with more accuracy and robustness than previous neural network and analytical methods. The unique algorithm suggested here converges to solutions by recursively solving an unsupervised learner (S[barbelow]elf-O[barbelow]rganizing F[barbelow]eature M[barbelow]ap) and supervised learner (B[barbelow]ack P[barbelow]ropagation). The SOFM manages the mapping of seafloor bottom types from modeled hyperspectral training data. The spatial output provided by the SOFM and the hyperspectral remotely sensed spectra from an image are then used by the BP to solve for water column depth. Previous studies are limited by their expression for the relationship between multiple seafloor bottom types and depth and thus their robustness. The final product of this research will be a versatile software tool applicable to a class of nonlinear inverse problems exhibiting a recursive dependency on water depth and seafloor bottom type.