Application of a PLS-Augmented ANN Model for Retrieving Chlorophyll-a from Hyperspectral Data in Case 2 Waters of the Western Basin of Lake Erie

Abstract

We present results that demonstrate the utility of machine learning techniques that are based on partial least squares (PLS) and artificial neural networks (ANNs) for estimating low-moderate chlorophyll-a (chl-a) concentrations in the western basin of Lake Erie (WBLE). Previous ocean color studies have resulted in a large number of algorithms that are based on spectral indices to estimate water quality parameters (WQPs) such as chl-a concentration from remote sensing reflectance. However, these spectral index algorithms are based on reflectance features at specific wavelengths and do not take advantage of the wealth of spectral information that is contained in hyperspectral data, and are often not easily adaptable to waters with conditions that are different from those in the datasets that were used to originally calibrate the indices. Recently, there have been efforts to use machine learning techniques that are based on ANNs and PLS regression to exploit the spectral richness contained in hyperspectral data and retrieve WQPs. In this study, we have combined an ANN model with output from PLS regression to retrieve chl-a concentration from hyperspectral data in the WBLE. We compared the results from the PLS-ANN method to those that were obtained from a band-ratio algorithm that is based on reflectances in the blue and green spectral regions, a band ratio algorithm that is based on reflectances in the red and near-infrared (NIR) spectral regions, and a PLS-only approach. For a dataset that was collected in 2012, with chl-a concentrations ranging from 0.48 to 21.2 µg/L, the PLS-ANN method yielded a root mean square error (RMSE) of 1.22 µg/L, whereas the blue-green ratio algorithm yielded an RMSE of 1.75 µg/L, the NIR-red ratio algorithm yielded an RMSE of 1.95 µg/L, and the PLS-only approach yielded an RMSE of 1.95 µg/L. The PLS-ANN method takes advantage of the PLS regression to identify specific wavelengths that contain most information about the variation in chl-a concentration, minimize spectral collinearity and redundancy in the data, and simplify the neural network’s input structure. The better performance of the PLS-ANN method can also be attributed to the neural network’s ability to account for nonlinearity in the relationship between chl-a concentration and spectral reflectance. The results indicate that the PLS-ANN method can be reliably used to estimate and monitor low-moderate chl-a concentrations in optically complex waters.