Knowledge-based learning for classification of hyperspectral data

Access full-text files

Date

2007-05

Authors

Chen, Yang-Chi, 1973-

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

This research focuses on three critical issues related to land cover classification using hyperspectral data: i) robust classification of high dimensional input data; ii) utilization of contextual spatial information; and iii) knowledge transfer for classification of data for which little or no labeled samples are available. An integrated max-cut hierarchical decomposition algorithm that uses support vector machines to classify multi-class land cover data is proposed to address the high dimensional input problem. The hierarchical support vector machine (HSVM) classifier solves a series of max-cut binary set partitioning problems to hierarchically and recursively partition the set of classes into two subsets until pure leaf nodes are obtained. Support vector machines are used at each internal node of the hierarchy to construct the binary decision boundary. It is shown to perform well with limited amount of ground truth. Although hyperspectral data provide new capabilities for discriminating spectrally similar classes, it is often useful to incorporate reliable spatial information. A knowledge-based stacking approach is proposed to utilize spatial information within homogeneous regions and at class boundaries. The proposed max-cut HSVM approach (MC-HSVM) learns the location of the class boundary and combines original bands with the extracted spectral information of a neighborhood to train the HSVM classifier. An ensemble of majority filtering and MC-HSVM is also investigated to handle complex spatial neighborhoods through a switch process. Since the spectral signatures could be affected by many uncontrollable factors, a classifier must capture the resulting variations in spectral signatures. Inspired by nonlinear manifold learning, a shortest path k-nearest neighbor classifier (SkNN) is proposed for the analysis of spatially disjoint data and multi-temporal images. The ability to update an existing model so that it performs well on images with no labeled data leads to many potential applications of land cover classification. As a result, this research simplifies the land cover classification process and increases the accessibility of hyperspectral sensors through the development of intelligent classification algorithms. Algorithms proposed in this research help solve the three critical problems outlined previously and achieve the objective of this study: to develop efficient, knowledge-based classification procedures for hyperspectral sensed image data.

Description

text

Citation