Artificial neural networks for modeling the transfer function between marine reflectance and phytoplankton pigment concentration

Abstract

A neural network methodology is developed to estimate the near-surface phytoplankton pigment concentration of case I waters from spectral marine reflectance measurements (ocean color) at the Sea-viewing Wide Field-of-view Sensor (SeaWiFS) visible wavelengths. The advantages of neural network approximation, i.e., association of nonlinear complexity, smoothness, and reduced sensitivity to noise, are demonstrated. When applied to in situ California Cooperative Oceanic Fisheries Investigations data, the neural network algorithm performs better than the reflectance ratio algorithms. Relative rms errors on pigment concentration are reduced from 61 and 62 to 38%, and absolute rms errors are reduced from 4.43 and 3.52 to 0.83 mg m-3. When applied to SeaWiFS-derived imagery, there is statistical evidence that the neural network algorithm filters residual atmospheric correction errors more efficiently than the standard SeaWiFS bio-optical algorithm. Copyright 2000 by the American Geophysical Union.