Discrimination of Seagrass Species and Cover Classes with in situ Hyperspectral Data

Pu, R.; Bell, S.; Baggett, L.; Meyer, C., and Zhao, Y., 2012. Discrimination of seagrass species and cover classes with in situ hyperspectral data. Journal of Coastal Research, 28(6), 1330-1344. Coconut Creek (Florida), ISSN 0749-0208. Seagrass habitats support a variety of ecosystem functions and an increasing interest has emerged for utilizing remote sensing to acquire information on the spatial extent and abundance of seagrass vegetation. Here we report on hyperspectral data collected from a combined laboratory and field-based study to examine the spectral qualities of seagrass species and evaluate whether seagrass species and levels of seagrass cover could be distinguished using true in situ hyperspectral data collected by a spectrometer overlying sea-grass-dominated vegetation in a shallow water setting in the central west coast of Florida. We also analyzed hyperspectral data measured in the lab to compare with those from in situ collections. Using a set of 97 field measurements we compared spectra qualities for different seagrass species, levels of seagrass cover, water depths, and substrate types over wavelengths 400-800 nm, using spectral data preprocessing and data transformation. Optimal wavelengths for identifying seagrass species and levels of seagrass cover were determined by two-sample t-tests. We also utilized principal component analysis (PCA) on spectra to evaluate if a set of first five PCs could be used to discriminate effectively among seagrass species and levels of seagrass cover. The experimental results indicate that the best accuracies for identifying species were produced with the data set of the second -derivative normalized spectra. The optimal wavelengths were 450, 500, 520, 550, 600, 620, 680, and 700 nm, most of which are related to the peaks of reflectance and absorption bands by photosynthetic and accessory pigments. A set of five optimal bands produced higher accuracies for identifying seagrass species (overall accuracy = 73% and average accuracy = 75%) compared with those from use of PCA. Data preprocessing techniques were demonstrated to be effective for improving discriminant accuracies of species and levels of seagrass cover.