Mapping Tree Species in Coastal Portugal Using Statistically Segmented Principal Component Analysis and Other Methods

Abstract

Hyperspectral sensors record radiances in a large number of wavelengths of the electromagnetic spectrum and can be used to distinguish different tree species based on their characteristic reflectance signatures. Reflectance spectra were measured from airborne hyperspectral AISA Eagle/Hawk imagery in order to identify different Mediterranean tree species at a coastal test site in Portugal. A spectral range from 400 to 2450 nm was recorded at 2-m spatial resolution. The hyperspectral data are divided into five spectral data ranges. The chosen ranges for segmentation are based on statistical properties as well as on their wavelengths, as radiances of a particular wavelength may overlap with neighboring wavelengths. Principal component analysis (PCA) is applied individually to each spectral range. The first three principal components (PCs) of each range are chosen and are fused into a new data segment of reduced dimensionality. The resulting 15 PCs contain 99.42% of the information content of the original hyperspectral image. These PCs were used for a maximum likelihood classification (MLC). Spectral signatures were also analyzed for the hyperspectral data, and were validated with ground data collected in the field by a handheld spectro-radiometer. Different RGB combinations of PC bands of segmented PC image provide distinct feature identification. A comparison with other classification approaches (spectral angle mapper and MLC of the original hyperspectral imagery) shows that the MLC of the segmented PCA achieves the highest accuracy, due to its ability to reduce the Hughes phenomenon.