Principal component analysis applied to feature-oriented band ratios of hyperspectral data: a tool for vegetation studies
This paper presents a simple but effective method to identify and map the distribution of vegetation using hyperspectral data, which involves band ratios and principal component analysis (PCA). Using spectral data from the literature, we devised spectral indices focused on specific vegetation constituents (band positions are in nm): chlorophyll-a (461/422 and 807/638); chlorophyll-b (520/470 and 807/648); α-carotene (520/442); carotenoids (539/490 and 807/490); anthocyanin (510/530); phytochrome P730 (845/730), phytochrome P660 (778/658); lignin (1028/2101); cellulose (2211/2400); nitrogen (1731/1691); and leaf water content (1066/1452). In a second step, these indices were submitted to PCA. PCA eigenvector loadings were examined to decide which of the principal component (PC) images concentrate information directly related to the spectral signatures of specific vegetation constituents. An important aspect of this approach was that it predicted whether the target surface type would be highlighted by dark or bright pixels in the relevant PC image. Finally, the optimal PCs are selected and combined in a colour composite. In this study, we applied this method to atmospherically corrected Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) data collected in the Alto Paraíso de Goiás (APG) region of central Brazil. The results from identifying and mapping the distribution of vegetation in the APG region were very encouraging. Photosynthetic and non-photosynthetic vegetation or a mixture of both were mapped on the basis of feature-oriented PC loadings. Discriminations between vegetation types were made, including identifying riparian forests, burn grasslands and resurgence zones, crops and several types of savannah and pastures.