Adaptive multilevel classification and detection in multispectral images

A novel multilevel adaptive pixel classification and detection (AMLCD) method for detecting pixel and subpixel-size targets for multispectral images is presented. The AMLCD method takes into account both spectral and spatial characteristics of the data. In the first level of processing, the principal background end members are obtained using the K-means clustering method. Each pixel is examined next for classification using a minimum-distance classifier with the principal end members obtained in the previous level. In the second level, the neighborhood of each unclassified pixel is analyzed for inclusion of candidate end members in an unmixing procedure. If the list of candidate background classes is empty, the conditions for their inclusion are relaxed. The fractions of neighborhood and target signatures for the unclassified pixels are determined by means of a linear least-squares method in the third level. If the results of unmixing are not satisfactory, the list of candidate clusters is renewed. Target detection within each pixel is performed next. The last processing level determines the size and location of detected targets with a clustering analysis methodology. Target size and location are estimated on the basis of the sum and weighted vector mean, respectively, of the mixing fractions of the neighboring pixels. The AMLCD method was successfully applied to both synthetic and Airborne Visible/ Infrared Imaging Spectrometer (AVIRIS) hyperspectral imagery data sets.