Improving Classification of Remotely Sensed Data Using Best Band Selection Index and Cluster Labelling Algorithms
Teoh, Chin Chuang (2005) Improving Classification of Remotely Sensed Data Using Best Band Selection Index and Cluster Labelling Algorithms. PhD thesis, Universiti Putra Malaysia.
Methods for improving supervised and unsupervised classification of remotely sensed data were developed in this study. Supervised classification of remotely sensed data requires systematic collection of training samples for classes of interest. Image visual interpretation is important in training samples collection because it incorporates association information of surrounding pixels, such as texture and context, hence making the training samples collection process more easy and accurate. Once training samples for each class are collected, the training statistics for each class and band are extracted to select those bands, which are most effective in discriminating each class of information from all others for classification. In remote sensing application, deciding the best band combination for image visualization and classification is relatively difficult and time consuming. In addition, the best band selected for image classification is not necessarily the best for classification.A Best Band Selection Index (BBSI) algorithm was developed which is capable of selecting the best band combination for image visualization and supervised classification. This BBSI is calculated by two components, one based on class mean (or cluster mean) difference and the other based on correlation coefficients. Using Landsat Thematic Mapper (TM) and ModisIAster Airborne Simulator (hMSTER) images as the test datasets, the BBSI algorithm was compared to the Optimum Index Factor (OIF) algorithm in selection of the best three-band combination for image visualization. The comparison results between BBSI and OIF indicated that, both algorithms correctly predicted the best three-band combination that provided useful information for image visualization in the Landsat TM dataset. However, both algorithms tested on MASTER dataset produced different results. The image quality of band combination selected by BBSI was smoother and better than OIF. The BBSI was also compared to the Jefieys-Matusita distance (JM-distance) algorithm in selection of the best four-band combination for supervised classification of Landsat TM and MASTER datasets. The comparison results between BBSI and JM-distance showed that, both algorithms accurately selected the best four-band combination that yielded the highest overall accuracy classification map with value of 91% in the Landsat TM dataset. Meanwhile, the comparison results in the MASTER dataset showed that, the overall accuracy classification map for band combination selected by BBSI with value of 89.7% was slightly higher than band combination selected by JM-distance with value of 89.2%.Umpervised classification of remotely sensed data consists of cluster generation and cluster labelling steps. A method was developed to improve the cluster generation and clusters labelling processes in unsupervised classification of the Landsat TM and MASTER datasets. In cluster generating process, the developed BBSI algorithm was used to select the best band combination for generating cluster by using Iterative self- Organizing Data Analysis (ISODATA) technique. The cluster generation results showed that, the BBSI accurately selected the best four-band combination generating very low mixed classes of clusters. In cluster labelling process, a cluster labelling algorithm based on calculation of minimum-distance (MD) between cluster mean and class mean was developed to label the clusters. This algorithm was compared to co-spectral plot method for labelling clusters the clusters generated in Landsat TM dataset. The comparison results show that, the clusters labelled by the cluster labelling algorithm were the same as using co-spectral plot. The cluster labelling algorithm was also compared to maximum-likelihood supervised classifier in the production of classification map for MASTER dataset. The comparison showed that, the accuracy of the unsupervised classification map with value of 88.4% that was generated by using the cluster labelling algorithm was slightly more than the maximum-likelihood supervised classification map with value of 87.5%. The advantage of the cluster labelling algorithm compared to co-spectral plot and maximum-likelihood classifier was the algorithm provided a rapid production of high accuracy classification map.
Repository Staff Only: Edit item detail