Remote Sensing of the Environment : Rio Grande Delta Marshes
Classification Methodology
High resolution hyperspectral data from two airborne sensors were acquired at several locations in the Lower Rio Grande Valley, Texas on two occasions. In April 1999, CASI (compact airborne spectrographic imager) collected 17 bands of data over the Palo Alto Battlefield, Tio Cano Marsh, and Stover Point. The CASI instrument collects data in specified spectral ranges in the visible and near infrared portions of the spectrum. On September 1999, high resolution HYMAP data were acquired over Tio Cano Marsh, Stover Point, and Santa Ana National Wildlife Refuge. The HYMAP instrument collects 128 channels from 380nm - 2500nm. While the HYMAP data collects more spectral information and it highly calibrated, it is substantially more expensive for data collection. The spatial resolution of both instruments are approximately 4m.
The two classification algorithms developed by CSR which were applied to the hyperspectral data sets are; Hierarchical Tree Classifier (HTC) and Pairwise classifier with Best Basis feature extraction. The HTC recursively decomposes a C-class classification problem into C-1 two-class problems, where C is the number of classes. A framework is utilized to partition classes into groups depending on their similarities. Hence, classes which are similar to each other, e.g. different marsh types, trees and bushes, are considered separately during the classification process. The second algorithm selects bands (or features) that provide the best possible discrimination between each pair of the C-1 class problem. Once bands are selected the ones which are adjacent and highly correlated are grouped where a weight is assigned to each one of them representing its contribution to the discrimination of each class pair. Accordingly, the best discriminatory bases are obtained (one for each class pair) and the pairwise classification technique is used at each pixel which determines the most probable class for it given all classes. Moreover, by summing up the individual weights assigned to a band for each class pair discrimination problem, we can obtain a measure of its overall contribution (importance) to the resulting classification. The followings give the results of these two algorithms for each data set.
What is interesting to note is that for both classifications which used the HYMAP data, the bands with wavelengths higher than 1.5 µm were not selected as having importance for either classification. For all the classifications, the Hierarchical Tree classifier seemed to perform the best, in terms of classification accuracies as well as the actual thematic result.
Classification Results
|
Last Modified: Wed Apr 14, 1999
CSR/TSGC Team Web
|
