IASPEI - Picture Gallery

IUGG XXIV General Assembly July 2-13, 2007 Perugia, Italy
(S) - IASPEI - International Association of Seismology and Physics of the Earth's
Interior
JSS006 Poster presentation 1951
Ground Based Imaging Spectroscopy and Supervised Classification
algorithms. New tools for the 21st century earthquake geologist.
Dr. Daniel Ragona
Geological Sciences San Diego State University
Tom Rockwell, Bernard Minster
We present a new methodology to assist the description, interpretation and archival of stratigraphic and
structural information from field exposures. Portable hyperspectral scanners were used in the field or
lab to collect high-quality spectroscopic information at high spatial resolution (pixel size ~ 0.5 mm at 50
cm) over frequencies ranging from visible to short wave infrared (VNIR-SWIR). A variety of robust
algorithms generated fast and accurate mapping of the images supplying the geologist with new
information to facilitate the interpretation process. This methodology, named hereinafter Ground-Based
Imaging Spectroscopy (GBIS), provide a new set of tools that complement the traditional techniques of
geological analysis. The principal benefits can be categorized in three groups. 1) Acquisition of a new
type of data. High-resolution VNIR-SWIR hyperspectral datasets provide textural and mineralogical
information at each pixel of the image, in some cases invisible to the human eye. 2) Quantitative
analysis of the spectra using robust algorithms allows fast classification of the images generating
alternative ways of mapping and correlation. 3) Data sharing and archival. Hyperspectral datasets
constitute an objective description of the geological materials that can be digitally transmitted to a
world-wide base of colleges for further analysis and interpretation. It is also an ideal format to construct
GIS-type data bases of geological exposures, samples and cores. These advantages can benefit
earthquake geology studies that require objective high-resolution geological description to resolve the
earthquake history at a site. To evaluate the methodology we acquired high-spatial resolution spectral
data of a large sample (60 x 60 cm) and four cores of faulted sediments from a paleoseismic excavation
site using portable push broom Specim hyperspectral scanners. These data, which contains hundreds of
narrow contiguous spectral bands between 400 and 2403 nm, were processed to obtain the reflectance
spectra at each pixel. The dataset was analyzed using traditional spectroscopic methods and processed
in a variety of ways to enhanced stratigraphic and structural relationships not obvious to the human
eye. Additionally we used a neural network algorithm (MLP) to generate classification models of eight
different types of materials (classes) observed on the samples. The best models were applied to the
hyperspectral dataset to obtain detailed and accurate classification maps of the samples. The results of
classification show that hyperspectral images classified with supervised algorithms can be used to
properly map sediments, even those of very similar compositions and grain sizes. In conclusion,
Ground-Based Imaging Spectroscopy is a very useful complement to traditional geological description
techniques. It not only provides new ways to visualize the geological data but also can be used as a
quantitative tool for fast and objective classification of geological materials in field or lab settings.
Keywords: paleoseismology, hyperspectral, neuralnetworks