| Retrieving inherent optical properties of water from remote sensing multispectral reflectance measurements is difficult due to both the complex nature of the forward modeling and the inherent nonlinearity of the inverse problem. In such cases, neural network (NN) techniques have a long history in inverting complex nonlinear systems. In this study we present the construction and validation of three NN's working in parallel to model the inverse problem for both case 1 and case 2 waters. The first NN is used to relate the remote sensing reflectance at available MODIS visible wavelengths (except the 678 nm fluorescence channel) to the absorption and backscatter coefficients at 442nm (peak of phytoplankton absorption). The second NN separates algal and non-algal absorption components, outputting the ratio of algal to non-algal absorption and the third, in a similar manner, outputs the ratio of non-algal particulate to dissolved absorption coefficient. With the outputs of these statistically derived networks we can thereafter analytically obtain the absorbing properties of the three known major water components. These include the color dissolved organic matter (CDOM), phytoplankton, and non-algal particulates (NAP). The resulting synthetically trained algorithm is tested using both the NASA Bio-Optical Marine Algorithm Data set (NOMAD), as well as our own field data sets from the Chesapeake Bay and Long Island Sound, New York. Very good agreement is obtained, when the retrievals are compared with the measurements of both the NOMAD dataset as well as our field data. Furthermore we apply our algorithm on satellite imagery and finally we test to what extent the empirical relationships used to describe the IOPs can be applied. |
