Page Banner

United States Department of Agriculture

Agricultural Research Service

Title: Using Casi Hyperspectral Imagery to Detect Mortality and Vegetation Stress Associated with a New Hardwood Forest Disease

Authors
item Ruiliang, Pu - UNIVERSITY OF CA-BERKELEY
item Kelly, Maggi - UNIVERSITY OF CA-BERKELEY
item Anderson, Gerald
item Gong, Peng - UNIVERSITY OF CA-BERKELEY

Submitted to: Journal of Photogrammetric Engineering and Remote Sensing
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: April 25, 2006
Publication Date: N/A

Interpretive Summary: Sudden Oak Death Syndrome is killing thousands of oak and tanoak trees along the coast of California. Several methods have been evaluated to assess and monitor the problem. This paper evaluated an imaging system capable of acquiring 48 bands of information with a 2 m pixel. The experimental results indicated that the more detailed classification level used in the multilevel classification scheme lead in general to higher classification accuracy. We also found that principal components derived from visible and NIR bands yielded more accurate results than did the other feature extraction methods, especially for the classification scheme with only vegetation classes and for the scheme highlighting vegetation stress. Stressed trees could be separated from other oaks (albeit with poor accuracy levels) only when there were the only two classes being evaluated. In general, the spectral resolution of hyperspectral data (in our case, the CASI data had 11 nm spectral resolution) provided better classification accuracies than that of multispectral data (we simulated multispectral resolutions by averaging the CASI data bandwidth of 11 nm over 55 nm).

Technical Abstract: A new disease affecting hardwood forests in California, called sudden oak death (SOD), has resulted in hundreds of thousands of dead oak (Quercus) and tanoak (Lithocarpus) trees in the state's coastal forests. Monitoring mortality associated with the disease and identifying stressed trees is a priority for management and regulation of the disease. Multispectral remote sensing methods have proved to be successful in mapping and monitoring forest health and distribution; we evaluated the use of an airborne hyperspectral imaging sensor, the Compact Airborne Spectrographic Imager (CASI), for mapping mortality and detecting tree stress. Hyperspectral imagery is rich in spectral data, and this fact necessitates considerable thought be devoted to the pre-processing and classification methods employed. Because certain features are spectrally similar, we first developed a multilevel classification scheme to increase classification accuracy of final classes. We next transformed the CASI radiance values to reflectance, and corrected each image for topography. Using these four datasets (radiance, reflectance, corrected radiance and corrected reflectance), we developed four feature extraction methods of ten features each, and then classified each of the 16 datasets using a maximum likelihood classifier, and tested the relative accuracies of each preprocessing and feature extraction method across classification scheme. The experimental results indicated that the more detailed classification level used in the multilevel classification scheme lead in general to higher classification accuracy. We also found that principal components derived from visible and NIR bands yielded more accurate results than did the other feature extraction methods, especially for the classification scheme with only vegetation classes and for the scheme highlighting vegetation stress. Stressed trees could be separated from other oaks (albeit with poor accuracy levels) only when there were the only two classes being evaluated. In general, the spectral resolution of hyperspectral data (in our case, the CASI data had 11 nm spectral resolution) provided better classification accuracies than that of multispectral data (we simulated multispectral resolutions by averaging the CASI data bandwidth of 11 nm over 55 nm). We also found that topographic corrected CASI data calibration did not lead to a classification accuracy improvement in this analysis.

Last Modified: 4/23/2014
Footer Content Back to Top of Page