|Miao, Xin - UNIV OF BERKELEY, CA|
|Pu, Ruiliang - UNIV OF BERKELEY, CA|
|Gong, Peng - UNIV OF BERKELEY, CA|
Submitted to: Photogrammetric Engineering and Remote Sensing
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: November 1, 2005
Publication Date: September 1, 2007
Citation: Miao, X., Gong, P., Pu, R., Carruthers, R., Heaton, J. 2007. Detection of Yellow Starthistle Through Band Selection and Feature Extraction from Hyperspectral I magery. Photogrammetric Engineering & Remote Sensing 73(9): 10051015 Interpretive Summary: Yellow starthistle is one of the worst invasive weeds to infest California and other western states. It grows in rangelands, pastures, natural areas, along roadsides and other disturbed habitats, infesting approximately 15 million acres. This weed produces toxins that cause brain lesion and death in horses, is a poor forage plant for other wildlife and livestock due to its spiny flower heads, it uses excessive amounts of valuable water out competing native species and other beneficial plants, and induces wildfires. A combined effort between UC Berkeley and the USDA-ARS has allowed yellow starthistle plants to be detected using aerial borne hyperspectral image capture and evaluation. Spectral analysis of reflectance data across a wide range of wavelength led to an optimal selection of patterning that allowed this invasive species to be characterized and separated from other background vegetation. This is the first step in allowing remote sensed data to be used to aid land managers in controlling this noxious and invasive weed.
Technical Abstract: To effectively display hyperspectral imagery for visualization purposes, the three RGB channels should be selected or extracted from a hyperspectral image under the criteria of maximum information or maximum between-class separability. Seven band selection (OIF, SI, CI, divergence, transformed divergence, B-distance, JM-distance) and five feature extraction (principal component analysis, linear discriminant analysis, class-based PCA, segmented PCT (SPCT), independent component analysis) methods and their variations are examined and compared using CASI hyperspectral imagery with the goal of detecting Centaurea solstitialis (yellow starthistle, YST), an invasive weed, in an annual grassland in California. Three indicators, information index (Infodex), separability index (Sepadex) and average correlation coefficient (ACC) are proposed to evaluate the quality of the generated images. The results suggest that both the combination of three SPCT channels and the combination of the second PCA channel with the positive and negative of the first LDA channels (PCA2, LDA1, -LDA1) can enhance our ability to visualize the distribution of YST in contrast to the surrounding vegetation. SPCT is also a good alternative to the conventional band selection methods in our study.