Showing papers by "David A. Landgrebe published in 1994"

The effect of unlabeled samples in reducing the small sample size problem and mitigating the Hughes phenomenon

Abstract: The authors study the use of unlabeled samples in reducing the problem of small training sample size that can severely affect the recognition rate of classifiers when the dimensionality of the multispectral data is high The authors show that by using additional unlabeled samples that are available at no extra cost, the performance may be improved, and therefore the Hughes phenomenon can be mitigated Furthermore, by experiments, they show that by using additional unlabeled samples more representative estimates can be obtained They also propose a semiparametric method for incorporating the training (ie, labeled) and unlabeled samples simultaneously into the parameter estimation process >

The Effect of Unlabeled Samples in Reducing the Small Sample Size Problem and Mitigating the

Abstract: In this paper, we study the use of unlabeled sam- ples in reducing the problem of small training sample size that can severely affect the recognition rate of classifiers when the dimensionality of the multispectral data is high. We show that by using additional unlabeled samples that are available at no extra cost, the performance may be improved, and therefore the Hughes phenomenon can be mitigated. Furthermore, by ex- periments, we show that by using additional unlabeled samples more representative estimates can be obtained. We also pro- pose a semiparametric method for incorporating the training (Le., labeled) and unlabeled samples simultaneously into the parameter estimation process.

High dimensional feature reduction via projection pursuit

Abstract: The recent development of more sophisticated remote sensing systems enables the measurement of radiation in many more spectral intervals than previously possible. An example of that technology is the AVIRIS system, which collects image data in 220 bands. As a result of this, new algorithms must be developed in order to analyze the more complex data effectively. Data in a high dimensional space presents a substantial challenge, since intuitive concepts valid in a 2-3 dimensional space do not necessarily apply in higher dimensional spaces. For example, high dimensional space is mostly empty. This results from the concentration of data in the corners of hypercubes. Other examples may be cited. Such observations suggest the need to project data to a subspace of a much lower dimension on a problem specific basis in such a manner that information is not lost. Projection pursuit is a technique that will accomplish such a goal. Since it processes data in lower dimensions, it should avoid many of the difficulties of high dimensional spaces. The authors investigate some of the properties of projection pursuit. >

Fast Parzen density estimation using clustering-based branch and bound

Abstract: This correspondence proposes a fast Parzen density estimation algorithm that would be especially useful in nonparametric discriminant analysis problems. By preclustering the data and applying a simple branch and bound procedure to the clusters, significant numbers of data samples that would contribute little to the density estimate can be excluded without detriment to actual evaluation via the kernel functions. This technique is especially helpful in the multivariant case, and does not require a uniform sampling grid. The proposed algorithm may also be used in conjunction with the data reduction technique of Fukunaga and Hayes (1989) to further reduce the computational load. Experimental results are presented to verify the effectiveness of this algorithm. >

Effect of radiance-to-reflectance transformation and atmosphere removal on maximum likelihood classification accuracy of high-dimensional remote sensing data

Abstract: Many analysis algorithms for high-dimensional remote sensing data require that the remotely sensed radiance spectra be transformed to approximate reflectance to allow comparison with a library of laboratory reflectance spectra. In maximum likelihood classification, however, the remotely sensed spectra are compared to training samples, thus a transformation to reflectance may or may not be helpful. The effect of several radiance-to-reflectance transformations on maximum likelihood classification accuracy is investigated. The authors show that the empirical line approach, LOWTRAN7, flat-field correction, single spectrum method, and internal average reflectance are all non-singular affine transformations, and that non-singular affine transformations have no effect on discriminant analysis feature extraction and maximum likelihood classification accuracy. (An affine transformation is a linear transformation with an optional offset.) Since the Atmosphere Removal Program (ATREM) and the log residue method are not affine transformations, experiments with Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) data were conducted to determine the effect of these transformations on maximum likelihood classification accuracy. The average classification accuracy of the data transformed by ATREM and the log residue method was slightly less than the accuracy of the original radiance data. Since the radiance-to-reflectance transformations allow direct comparison of remotely sensed spectra with laboratory reflectance spectra, they can be quite useful in labeling the training samples required by maximum likelihood classification, but these transformations have only a slight effect or no effect at all on discriminant analysis and maximum likelihood classification accuracy. >