HR: 15:25h
AN: NG33B-07    [Abstracts]
TI: Detection of Local Anomalies in High Resolution Hyperspectral Imagery Using Geostatistical Filtering and Local Spatial Statistics
AU: * Goovaerts, P
EM: goovaerts@biomedware.com
AF: Biomedware Inc., 516 North State Street, Ann Arbor, MI 48104 United States
AU: Jacquez, G M
EM: jacquez@biomedware.com
AF: Biomedware Inc., 516 North State Street, Ann Arbor, MI 48104 United States
AU: Marcus, A W
EM: marcus@darkwing.uoregon.edu
AF: University of Oregon, Department of Geography, Eugene, OR 97403-1251 United States
AB: Spatial data are periodically collected and processed to monitor, analyze and interpret developments in our changing environment. Remote sensing is a modern way of data collecting and has seen an enormous growth since launching of modern satellites and development of airborne sensors. In particular, the recent availability of high spatial resolution hyperspectral imagery (spatial resolution of less than 5 meters and including data collected over 64 or more bands of electromagnetic radiation for each pixel offers a great potential to significantly enhance environmental mapping and our ability to model spatial systems. High spatial resolution imagery contains a remarkable quantity of information that could be used to analyze spatial breaks (boundaries), areas of similarity (clusters), and spatial autocorrelation (associations) across the landscape. This paper addresses the specific issue of soil disturbance detection, which could indicate the presence of land mines or recent movements of troop and heavy equipment. A challenge presented by soil detection is to retain the measurement of fine-scale features (i.e. mineral soil changes, organic content changes, vegetation disturbance related changes, aspect changes) while still covering proportionally large spatial areas. An additional difficulty is that no ground data might be available for the calibration of spectral signatures, and little might be known about the size of patches of disturbed soils to be detected. This paper describes a new technique for automatic target detection which capitalizes on both spatial and across spectral bands correlation, does not require any a priori information on the target spectral signature but does not allow discrimination between targets. This approach involves successively a multivariate statistical analysis (principal component analysis) of all spectral bands, a geostatistical filtering of noise and regional background in the first principal components using factorial kriging, and finally the computation of a local indicator of spatial autocorrelation to detect local clusters of high or low reflectance values as well as anomalies. The approach is illustrated using one meter resolution data collected in Yellowstone National Park. Ground validation data demonstrate the ability of the filtering procedure to reduce the proportion of false alarms, and its robustness under low signal to noise ratios. In almost all scenarios, the proposed approach outperforms traditional anomaly detectors (i.e. RXD) and fewer false alarms were obtained when using statistic S2 (average absolute deviation of p-values from 0.5 through all spectral bands) to summarize information across bands. Image degradation through addition of noise or reduction of spectral resolution tends to blur the detection of anomalies, leading to more false alarms, in particular for the identification of the least pure pixels. Results from the tailings site demonstrated that the approach still performs reasonably well for highly complex landscape with multiple targets of various sizes and shapes. By leveraging both spectral and spatial information, the technique requires little or no input from the user, and hence can be readily automated.
UR: http://www-personal.engin.umich.edu/$\sim$goovaert/IEEE03_P-27.pdf
DE: 6974 Signal processing
DE: 3360 Remote sensing
DE: 1869 Stochastic processes
DE: 0910 Data processing
SC: Nonlinear Geophysics [NG]
MN: 2004 AGU Fall Meeting