Feature Selection Methodologies Using Simulated Thematic Mapper Data

This study* investigated computer classification performances for forest and other cover types using Thematic Mapper Simulator (TMS) data collected by NASA's NSOOI scanner. Specifically, results based on the use of a common feature selection measure -transformed divergence (TD) were compared to those based on a principal component transformation for the purpose of evaluating the capabilities of each technique to define: (1) the optimum dimensionality for data sets of this type, and (2) the relative significance of the various wavelength bands with respect to their ability to discriminate among the various cover classes. Expected classification performances as indicated by a minimum Transformed Divergence (TDmin) criteria were compared to actual test classification results. The eigenvectors (i.e. principal components) and eigenvalues for both the overall and the individual class statistics used to classify the TMS data were also used to select waveband subsets to compare to the results from the subsets defined by TD(min). The results indicated that the use of four wavelength bands will produce considerably better classification than the use of only two or three wavelength bands. However, when more than four wavelength bands were used, overall and individual class performances increased only slightly, thereby indicating that an appropriate set of four wavelength bands probably provide the 'optimum' dimensionality. Classifications using various four wavelength band combinations showed the individual cover class preferences for certain wavebands. These preferences of * This work was supported by NASA under Contract No. NAS9-15889. both individual cover classes and of all classes combined were better indicated by a principal component analysis of the data than by a Transformed Divergence criteria. Further, the results support the use of eigenvectors for identifying the optimal or 'intrinsic' dimensionality of data sets of this type.