Automated training sample labeling using laboratory spectra

This paper presents a method for automatically labeling training samples in mineral identification problems. A previous work showed that an experienced human operator can successfully identify training samples by visually comparing the strong absorption features of the laboratory spectra to those of the adjusted remotely sensed spectra. However, it is obviously a time-consuming process. The purpose of this research is to automate the labeling process. The method proposed in this paper consists of a data preprocessor and correlation detection. This data preprocessor is used for shape adjustment and scale unification. It includes three parts: (1) the log residue method; (2) reference level setting; (3) signal normalization. The desired number of training samples can be specified to guarantee that enough training samples are drawn from the data set. After training samples are labeled, a series of statistical pattern recognition analysis methods are applied to the original remotely sensed data set to do classification. The data set used in the paper is Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) data taken over the Cuprite mining District, Nevada in 1992. The result is compared to a previous work that requires a human operator to visually label training samples and a rough geologic ground truth map. It shows that this automated labeling method not only saves human laboring but also achieves a fair performance of classification