Neural Network Based Retrieval of Inherent Optical Properties (IOPs) Of Coastal Waters of Oceans

Abstract

Inherent optical properties (IOPs) of the coastal oceans are modulated independently by the in-water optical constituents, which cause variations in the water leaving radiances or re-mote sensing reflectances. Accurate determination of IOPs and the optical constituents from water-leaving radiances or reflectances using conventional empirical ratio approaches fail in the coastal oceans. Alternate non-parametric approaches such as neural network (NN) based approaches can be developed to derive the parameters of interest using training datasets. Further NN based approaches can deal with the non-linearity of functional dependence between optical constituents and the IOPs.To retrieve the IOPs, earlier NN models used Levenberg-Marquardt with Bayesian Regularization as an optimizer for learning the weights of the model, which has a slow learning rate. Moreover, with low-resource datasets while retrieving IOPs till the third level, the probability of error propagation becomes high. To overcome these two problems, we present a Modified Neural Network (MNN) algorithm (modification of NN model [3] to retrieve Inherent Optical Properties (IOPs) of ocean waters, in which three Neural Networks (NN) were developed in parallel. Our method is based on the approach where we use the Adam optimizer, instead of the Levenberg-Marquardt since it has a faster training time. Also, the error propagation is observed to be very less even with low-resource data while retrieving IOPs at the third level, with a decent R² score.Results of the MNN algorithm indicate that MNN retrieves IOPs with an R² = 0.99 between measured and predicted values for b<inf>bp</inf>(443) and R² = 0.99 for a<inf>pg</inf>(443) at Level 1. Level-2 products give R² = 0.98 and R² = 0.99 between measured and predicted values for a<inf>pg</inf>(443) and a<inf>dg</inf>(443) respectively. Similarly Level-3 products give R² = 0.97 and R² = 0.51 between measured and predicted values for a<inf>g</inf>(443) and a<inf>d</inf>(443) respectively. The algorithm retrieves better R² score for all parameters except a<inf>d</inf>(443) compared to Semi Analytical Algorithm, Quasi Analytical Algorithm and NN algorithm by [3]. The new technique has the advantage of faster convergence and better generalization capacity for deriving IOPs from complex waters. The new algorithm is also able to separate gelbstoff and detrital absorption.