Evaluating the performance of artificial neural network techniques for pigment retrieval from ocean color in Case I waters

In the present paper, we report on a method to retrieve the pigment concentration in Case I waters from ocean color. The method is derived from radiative transfer (RT) simulations and subsequent application of artificial neural network (ANN) techniques. Information on absorption and total scattering of pure seawater, colored dissolved organic matter, and marine particles are mostly taken from published measurements or parameterizations. Additionally, a new model relating the backscattering of marine particles to pigment concentration and wavelength is introduced. The such defined inherent optical properties are input to a RT code in order to generate a synthetic data set of remote sensing reflectance spectra. This synthetic data set is then used for the training of a set of ANNs with the aim to approximate the functional relationship between ocean color and pigment concentration. The different ANNs are obtained by systematic variations of input parameters, architecture, and noise level added to the training data. The performance of each individual ANN‐based pigment retrieval scheme is assessed by applying it to the remote sensing reflectance spectra contained in the Sea‐viewing Wide Field‐of‐view Sensor (SeaWiFS) Bio‐optical Algorithm MiniWorkshop (SeaBAM) data set and comparing the retrieved pigment concentrations to those actually measured. The most successful ANN compares favorably with commonly used empirical pigment retrieval schemes. Compared, e.g., to the SeaWiFS algorithms OC2B and OC4, the square of the correlation coefficient r2 is increased from 0.924 (OC2B), respectively, 0.928 (OC4) to 0.934 (ANN). The root mean square error of the retrieved log‐transformed pigment concentration drops from 0.156 for OC2B, respectively, 0.151 for OC4 to 0.148 for the ANN‐based pigment retrieval scheme. Furthermore, the latter shows a higher resistance against noisy input data.