TOPIC: Processing and Integration, exploitation tools and
techniques.
TITLE: "Object Detection Utilizing a Linear Retrieval Algorithm for
Thermal Infrared Imagery"
M S Ramsey (Dept. of Geology, Arizona State Univ., Box 871404, Tempe, AZ 85287-1404;
ph. 602-965-1790; email: ramsey@elwood.la.asu.edu)
Problem:
Thermal infrared (TIR) remote sensing has been proven to be an extremely
valuable tool for lithologic descrimination in arid environments. However,
it has suffered full implementation partially due to the lack of long term
instrument development with concentration favoring the visible/near infrared
(VNIR) region. Improvements in sensor technology, reliable airborne
instruments, and the impending launch of the EOS-A in 1998, have resulted in
an expansion of thermal imaging of the earth's surface over the past decade.
However, the lower signal to noise, caused by a reduced amount of total
returned energy, often necessitates longer surface dwell times, larger pixels
and fewer spectral channels than typically present in VNIR.
The need for accurate processing, detection, and data reduction techniques
becomes apparent because of these constraints. The use of sub-pixel extraction
alogrothims allows not only for this detection, but also estimates the relative
endmember percentages. Translated to an image format, results produce abundance
maps of the mineral or feature of interest. This input endmember can either be
derived from the lab, a spectral library, or the image itself. The fitting
procedure produces a root-mean-squared (rms) error image that functions as an
accuracy test for a particular fit. Techniques such as these are not new and
have in fact been used for many years in remote sensing. This study however is
a comprehensive attempt to quantify the limits of the model and specifically
concentrate only on the thermal infrared.
Methodology:
The atmospheric window located between 8 and 12 microns coincides with the
primary vibrational features that occur in all silicate minerals. In addition,
there is a linear relation between the mineral's proportion and the strength of
its absorption features. This linearity provides the justification for the use
of the least-squares minimization routine. The combination of linear mixing and
the absporbtion band location, make TIR remote sensing excellent for geologic
mapping in arid regions.
In order to constrain possible complications as well as understand the
effects of different surfaces, the algorithm was applied to laboratory data;
eolian sediments (Kelso Dunes, CA); fluvial reworked impact alluvium (Meteor
Crater, AZ); and silicic lava domes (Medicine Lake, CA). Detectability limits,
residual errors and mixing patterns in each of these areas were examined.
Samples acquired from several of the field sites were analyzed both
petrographically and spectrascopically in order to confirm the alogorithm's
results. In addition, extensive field mapping was performed at these locations.
Results:
Results from all the test areas indicate that linear spectral unmixing
can produce accurate results to within 10 percent depending on the site
and endmember choice. Artifical mixtures in the laboratory could be
predicted to within 2 percent with rms errors as low as 1.0 x 10-6, while
image results had average deviations of 5-10 percent from the known
abundances with errors on the order of 0.01. In addition, surface textural
variations are also able to be detected. The degree of vesiculation throughout
the silicic domes, for example, indicate emplacement time, volume and viscosity
of the lava. Image endmembers, while proving much simpler to implement, are
themselves mixtures, and quantitative results based solely on their use are
nearly impossible. Complications such as vegetation, thermal shadowing and
particle size reduce the accuracy of the model and limit its applicability in
highly vegetated areas.
Conclusions:
Linear spectral unmixing when applied to thermal infrared data provides
a technique for mineral mapping and texture variations. Its use should be
constrained to arid regions where soil exposure and thermal flux are the
greatest. The ability to predict mineral or rock abundances to within several
percent becomes extremely useful in tracking sediment transport,
desertification and potential hazard assessment in remote volcanic regions.
------------------
Presented at: ERIM Second International Airborne Remote Sensing Conference
Date: 1996