PRINCIPAL COMPONENT ANALYSIS (PCA) FOR ESTIMATING CHLOROPHYLL CONCENTRATION USING FORWARD AND GENERALIZED REGRESSION NEURAL NETWORKS

Abstract

Proper estimates of chlorophyll (Chl) concentrations are essential for estimating primary productivity, biomass, and so forth. This research studied the application of two types of artificial neural networks along with a multiple linear regression (MLR) model for chlorophyll estimation. A feed-forward neural network (FFNN) using the Levenberg–Marquardt algorithm and a generalized regression neural network (GRNN) were applied to estimate the Chl concentration in consideration of 14 biological, physical, and meteorological parameters as independent input variables. In this study, principal components analysis (PCA) was used to form a smaller input number of uncorrelated variables from 14 potential input data. Principal components were obtained by projecting the multivariate data vectors on the space spanned by the eigenvectors. Applying PCA, eight different groups of input data were distinguished. The root mean squared error (RMSE), average of the absolute relative errors (AARE), and the correlation coefficients were applied as comparison criteria. Also, the residual analysis was examined in order to determine the operational behavior of each model in addition to the three previously mentioned performance criteria. The results of a case study (Hempstead East Marina, New York) showed that the estimated accuracy of chlorophyll concentration using neural networks was higher than the accuracy of chlorophyll estimation using the MLR model in the area. In general, most of the GRNN structures were superior to conventional FFNN structures. However, the best structure was found to be FFNN-8, which took advantage of using the entire set of input variables. A RMSE value of 4.15 was achieved by FFNN-8 in comparison with the values of 4.42 and 4.47 achieved by GRNN-8 and MLR-8, respectively. These results were confirmed by the analysis of the correlation coefficients between the measured and estimated Chl concentration on the one hand and the analysis of the AARE on the other. Residual analysis demonstrated that MLR and FFNN models overestimated the simulated Chl, whereas GRNN tended to underestimate the results.