Retrieval of Water Quality Parameters from NOAA Hyperspectral Data Using Neural Networks - Western Lake Erie

Anthropogenic eutrophication in Lake Erie, especially in the western basin, has adverse impacts on the ecosystem and economy in that region. As a result, it is very crucial to detect and solve the algal problem of Lake Erie. However, it is very expensive and exacting to sample and test the water quality of every location in Lake Erie. To tackle this challenge, remote sensing and machine learning techniques have advanced rapidly to detect concentration of optically active water constituents. In traditional remote sensing methods, estimation of these water quality parameters is performed by building linear regression using the ratio of (or individual) spectral bands. However, such regression predictor is not accurate and does not take the advantage of the whole shape of the spectral curve. Therefore, in this study, the hyperspectral curve shape is described by deriving the important peaks. In order to do this, the spectrum curves are first smoothed; and K-means clustering is applied to find the approximate range of important peaks. Next, the exact locations of the peaks are calculated in the smoothed curves for each range. Finally, because there is no significant differences between the original curves and the prominent-described curves, it is reasonable to take the wavelength, height, and the prominent of the most important peaks (i.e. the peak with the largest prominent) in each range as the input to develop a simple regression neural network. This study presents preliminary results to show the potential of using such approach to reconstruct hyperspectral data and derive the concentration of water quality parameters, including chl-a, phycocyanin, turbidity, dissolved and particulate matters, from NOAA-airborne data.