| for class
| derived
| evaluate
A COMPARISON OF CONVENTIONAL CLASSIFICATION METHODS AND A NEW | aed un
INDICATOR KRIGING BASED METHOD USING HIGH-SPECTRAL RESOLUTION | SAUCES
IMAGES (AVIRIS) | GEOL
| SURFA!
FREEK D. VAN DER MEER |
| The are:
International Institute for Aerospace Survey and Earth Sciences (ITC) | quatre
Department of Earth Resources Surveys | ere d
Research Scientist in Geological Survey | any
350 Boulevard 1945, P.O. Box 6, 7500 AA, Enschede, The Netherlands. | Y 1g
ISPRS Technical Commission VII | xposur
| extent \
metals v
ABSTRACT. High spectral resolution images (AVIRIS) providing detailed information on | The es
the surface mineralogy have been used to evaluate a new indicator kriging based | area Or
classification technique. This technique directly uses spectral information derived from of rhyol
AVIRIS data instead of information from training areas studied in the field. A small study | rocks ar
area of an imaging spectrometer data set covering the Cuprite mining district was selected opalizec
for its known occurrences of both kaolinite and alunite. Three "conventional" classification | contain!
methods were applied as well as the new indicator kriging based technique and results were | irregula
evaluated using shape characteristics of the classes and by comparison with local field | of the :
geologic information. Indicator kriging performed better than the conventional methods. | much a:
Furthermore, the new indicator kriging based method provides information on the reliability | soft, pc
of the classification which is lacking with the conventional methods. | i dd
| rocks, p
| to opal
Key Words: Image Classification, Indicator Kriging, AVIRIS, Cuprite Mining District, | kaolinit
Nevada | is treat
| E
INTRODUCTION the simultaneous collection of images in 224 contiguous | L
Remote sensing of earth's surface from aircraft and from
spacecraft provides information not easily acquired by
surface observations. Until recently, the main limitations of
remote sensing were that no subsurface information could be
obtained and that surface information lacked specification.
Conventional scanners (e.g. Landsat MSS and TM, and
SPOT) acquire information in a few separate spectral bands
of various widths, thus filtering to a large extent the
reflectance characteristics of the surface (Goetz & Rowan,
1981). Therefore, new scanner types were developed with
high spectral resolution yielding new image processing
techniques to cope with the increased amount of data. The
use of indicator kriging as classification routine is discussed
in this paper using high spectral resolution imagery although
the technique is also valid for conventional scanner data.
IMAGING SPECTROMETRY
The use of high spectral resolution remotely sensed imagery
for mineralogic mapping was first demonstrated in spectral
laboratory studies (e.g. Hunt, 1977). In 1981, airborne
spectrometer data were acquired using a sensor developed by
the GER corporation for one-dimensional profiling along a
flight line. The first imaging device was the Airborne
Imaging Spectrometer (AIS), developed at the Jet Propulsion
Laboratory. This instrument acquired data in 128 spectral
bands in the range of 1200-2400 nm with a field-of-view of
3.7 degrees (Vane & Goetz, 1985). In 1987 NASA began
data acquisition with an improved version of AIS called the
Airborne Visible/Infrared Imaging Spectrometer (AVIRIS;
see Macenka & Chrisp, 1987). This scanner makes possible
72
bands resulting in a complete reflectance spectrum for each
20*20 m. picture element (pixel) in the 400 to 2500nm
region with a sampling interval of 10 nm (Goetz et al.,
1985; Vane & Goetz, 1988; Porter & Enmark, 1987). The
field-of-view of the AVIRIS scanner is 30 degrees resulting
in a ground field-of-view of 10.5 km. The signal-to-noise
ratio is 100:1 at 700nm and 50:1 at 2200nm. The value of
this scanner lies in its ability to acquire a complete
reflectance spectrum for each pixel. Many surface materials
have diagnostic absorption features that are 20-40nm in
width (Hunt, 1979). Therefore, spectral imaging systems
which have 10nm wide bands can produce data with
sufficient resolution for resolving these features and
subsequent direct identification of those materials (Goetz,
1991). On the contrary, Landsat scanners, which have band
widths between 100 and 200nm cannot resolve these spectral
features. Analysis of high spectral resolution imagery for
mineral identification involves three steps: (1) the pre-
processing of the data to convert raw spectra into reflectance
spectra corrected for atmospheric influences, (2) extraction
of absorption features characterizing surface materials of
interest, and (3) evaluating for each pixel whether the
absorption feature is present or absent at the wavelength
(Okada et al., 1991).
This paper shows the potential use of indicator kriging based
techniques for image classification (the third processing step
mentioned above) of remotely sensed imagery in general and
high spectral resolution data in particular. AVIRIS data from
the Cuprite mining district were used to detect occurrences
of kaolinite and alunite based on their spectral
characteristics. Four bands defining the key absorption
features from these minerals are subsequently used as input