Absorption, reflection, scattering, refraction and other physical processes in the water column and atmosphere result in the spectral radiance observed by satellite sensors above the sea. Numerous models exist to solve the forward problem, i.e., given the constituents or optical properties of the ocean/atmosphere system, calculate the upwelling radiance at satellite altitudes. However, remote sensing of the ocean requires a solution of the inverse problem, i.e., given radiance observed from space, what are the constituents or optical properties of the water column? Nonlinearity and many-to-one properties prevent analytical solution of the inverse problem for optical remote sensing of the ocean. Neural networks are presented as a class of general nonlinear mappers from one vector space to another that provide empirical inversions which can lead to practical remote sensing algorithms even though the inverse problem cannot be solved analytically. The neural network approach is demonstrated for the retrieval of water depth, water clarity, and bottom type from hyperspectral imagery of coastal waters.

WHERE: TEC 340