John Cipar

Testing an automated unsupervised classification algorithm with diverse land covers

We test a new automatic unsupervised classification algorithm designed for hyperspectral images. The algorithm automatically determines the number of clusters in the image by finding dense regions of the pixel cloud. A variation on migrating means clustering is used to find the dense regions. Five scenes from an airborne AVIRIS data set are used to test the algorithm. The algorithm successfully finds the dominant land covers and many areally small land covers, such as roads and other man-made structures.

Background spectral library for Fort A.P. Hill, Virginia

We describe development of a background spectral library for Fort A. P. Hill, located in northeastern Virginia, based on hyperspectral images and an extensive land cover database. The database was used to identify areas of uniform land cover. The library contains means and standard deviations for 15 spectra measured in these uniform areas. Terrain categorization products consist of classification maps and fractional abundance maps determined by linear mixture analysis. There is excellent qualitative agreement between the linear unmixing results and the known land covers.

Statistical characterization of hyperspectral background clutter in the reflective spectral region

Hyperspectral imaging systems for daylight operation measure and analyze reflected and scattered radiation in p-spectral channels covering the reflective infrared region 0.4-2.5 microm. Consequently, the p-dimensional joint distribution of background clutter is required to design and evaluate optimum hyperspectral imaging processors. In this paper, we develop statistical models for the spectral variability of natural hyperspectral backgrounds using the class of elliptically contoured distributions. We demonstrate, using data from the NASA AVIRIS sensor, that models based on the multivariate t-elliptically contoured distribution capture with sufficient accuracy the statistical characteristics of natural hyperspectral backgrounds that are relevant to target detection applications.

Distinguishing between coniferous and deciduous forests using hyperspectral imagery

We test how well coniferous and deciduous forests can be distinguished using hyperspectral reflectance images. We find that within a given land cover, the average spectral angle and distance calculated from the mean spectrum ranges between 1-3 degrees and 0.2-0.4 du, respectively. These values are the lower limit on the ability of these metrics to separate land cover types. Our observations indicate that the major spectral difference between deciduous and coniferous forests is in the NIR plateau (750-1300 nm), a spectral region with high reflectance for green vegetation. Coniferous forests have NIR reflectances of approximately 0.2, while the reflectance for deciduous forests is 0.3 or greater. The shapes of the spectra, as measured by the spectral angle, are very similar. Individual pixel spectra, on the other hand, exhibit high variability, especially in the NIR plateau.

Distinguishing vegetation land covers using hyperspectral imagery

We use AVIRIS data collected at Fort A. P. Hill, Virginia, to evaluate how well airborne hyperspectral imagery can be used to distinguish vegetation land covers. Fort A. P. Hill is located in east-central Virginia and is heavily forested with a mix of deciduous and coniferous species native to the mid-Atlantic region. The location and extent of the forest species is documented in a land cover database compiled by the Fort for planning and resource protection purposes. The AVIRIS data set consists of several low-altitude (3.7-m GSD) flight lines on two dates: November 1999 and September 2001. Our goal is to characterize the both the natural variability of vegetation land covers using mathematical and biophysical metrics and to assess differences between land covers for classification purposes.

A Comparison of Forest Classification using Hyperion and AVIRIS Hyperspectral Imagery

We test how well a cluster-based unsupervised classification algorithm separates forest land covers. Our test data, Hyperion and AVIRIS images taken in northern Virginia during autumn, provide two spectrally distinct land covers: pine forests and senescent deciduous forests. We find that the algorithm successfully separates these land covers for AVIRIS data that has been spatially aggregated to simulate 30-m Hyperion GSD. The algorithm does not successfully separate the land covers for the Hyperion data.

Terrain categorization using a background spectral library

We describe development of a background spectral library for northeastern Virginia, USA, based on hyperspectral images and an extensive land cover database. The library consists of mean spectra and standard deviations measured in 14 areas of uniform land cover. Terrain categorization products include classification maps and fractional abundance maps determined by linear mixture analysis. There is excellent qualitative agreement between the linear unmixing results and the known land covers.

Summer to autumn changes in vegetation spectral indices of deciduous trees

The purpose of this work is to measure changes in deciduous tree reflectance spectra as a function of time from late summer to autumn senescence. Leaves were harvested from two maple trees growing in eastern Massachusetts. Reflectance in the 350-2500 nm range was measured in the laboratory on stacks of freshly-harvested leaves. We calculated a number of published spectral indices, finding that most of the indices varied remarkably little across the time period. In some case, the measurement uncertainty was small, although the measurements exhibited wide scatter over the time period. The Normalized Difference Vegetation Index showed only a slight downward drift throughout the measurement period. The red edge wavelength was observed to decrease dramatically from the summer growth period (~725 nm) to autumn senescence (~700 nm).

Unsupervised Land-Cover Classification Using a Cluster Algorithm

We describe a new approach to unsupervised classification that automatically finds dense parts of the hyperspectral data cloud. These dense regions are the cluster centers required for unsupervised classification. The approach is tested using AVIRIS data from central Texas

Hyperspectral signatures of an eastern North American temperate forest

We describe a new approach to unsupervised classification that automatically finds dense parts of the hyperspectral data cloud. These dense regions are identified as the cluster centers required for unsupervised classification. The approach is tested using AVIRIS hyperspectral imagery from central Texas that has spectrally well separated land covers. The algorithm is then applied to the more stressing case of separating coniferous and deciduous forests in eastern Virginia. We find that the major spectral difference is brighter reflectance in the NIR plateau for deciduous forests compared to adjacent coniferous stands. This difference is sufficient to distinguish the forest types, and is confirmed by comparison to ground truth information.