Inversion of inherent optical properties of water using artificial neural network techniques for coastal water

The efficient monitoring of coastal, or Case 2 waters by optical remote sensing has always been a challenging task. This study develops a neural network (NN) model to examine the possibility of accurate retrieval in waters where semianalytical and empirical algorithms do not perform satisfactory due to the large variability in the coexistence of particulates, dissolved matter and phytoplankton species. A multi-layer forward neural network was constructed to estimate the total absorption, phytoplankton absorption, total suspended matter and color dissolved organic matter absorption and total backscattering at the same time from in-situ measured water surface reflectance spectra. The neural network was trained using 60% of the 1000 reflectance spectra from a synthetic datasets that were generated using Hydrolight for water properties typical to coastal regions. Then, the neural network model was tested with the remaining 40% of the simulated reflectance spectra and applied to field data. Primarily the NN was trained and tested with the input of traditional visible channels. Thereafter one more channel was added from the UV region and the NN was again trained and tested. The retrievals with the addition of UV improve both in the simulated and field data..