Application of a partial least-squares regression model to retrieve chlorophyll-a concentrations in coastal waters using hyper-spectral data

Coastal and inland waters represent a diverse set of resources that support natural habitats and provide valuable ecosystem services to the human population. Monitoring the quality of these waters is essential to maintaining the resources they provide, and long-term monitoring may offer a better understanding of the relationship between human development and the health of these resource producers. The implementation of conventional monitoring is typically time-intensive and limited in geographic scale. Alternatively, the use of airborne and spaceborne remote sensors provides a synoptic view of water quality with better spatial coverage to more accurately identify dynamic and unique parameters. Concentrations of optically active constituents (OACs) such as suspended sediments and the phytoplankton pigment chlorophylla (CHLa), act as proxies for water quality and can be detected by optical sensors. Traditional remote sensing techniques were developed using multispectral sensors, and employ band ratio algorithms that seek to predict the concentrations of OACs in relation to water quality. In complex coastal waters, overlapping spectral signatures of OACs often confound these algorithms and reduce their predictive capacity. The objective of this study was to develop a dataset to test the predictive capabilities of partial least-squares regression, a multivariate statistical method, for hyperspectral remote sensing and in situ CHLa concentrations. This paper presents the model performance for a dataset developed in Long Bay, a ~160 km arcuate bay that spans the border between North and South Carolina. The model uses multivariate-based statistical modeling to capitalize on the spectral advantage gained by hyperspectral sensors when observing such waters. Following this approach, a multivariate-based monitoring tool for the prediction of CHLa concentrations is presented with a partial least-squares regression (PLSR) method using hyperspectral and laboratory-analyzed field data. The PLSR model was used based on its ability to identify wavelengths that are more sensitive to CHLa relative to other OACs. The model was able to explain 80% of the observed CHLa variability in Long Bay with root mean square error (RMSE) = 2.03 µg/L. This enhanced mode of water quality monitoring has potential to deliver insights to point-sources and problem areas that may contribute to a decline in the water quality of valuable inland and coastal systems.