| Abstract | In the field of remote sensing application, improving spatial resolution of satellite image data has increased within-class variances which is resulted in low classification accuracy of land-use categories. In effort to reduce this problem, most of current studies focus on an approach which creates texture features and incorporates this information to spectral features for classifying the whole image. This approach has two weaknesses as follows. Firstly, it could lead to an appearance of classification errors at the boundary between any two different categories. Secondly while improving the overall accuracy of the studied image, as compared with the classification approach based on spectral-values alone, it may reduce accuracy of certain categories. In addressing the above problems, this study develops a more systematic approach for using texture features in classification process. Instead of automatically creating texture information for the whole image, in the suggested approach texture features are created only for zones fulfilling the definition of confused categories. The texture information is then incorporated to spectral-values for re-classifying only these zones. The developed approach also introduces a mechanism to improve accuracies of categories whose accuracies are reduced due to adding texture information in the classification process. As a whole, comparing with the conventional approach, this more systematic approach of using texture features could maintain high classification accuracies of well-separated categories, increase the accuracies of confused categories and reduce or even eliminate errors at boundary areas. Experiment is carried out to illustrate the effectiveness of the suggested approach. Another problem encountered when using texture-features is about making decision in selection of a method for creating texture information. While there are a large number of experimentally developed methods for extracting texture-features, very few studies have been conducted to assess their powers in terms of classification accuracy and processing time. Hence, there is no tentative guideline about which extraction method should be used. As one of first steps to develop such a guideline, a comparative study based on experimental approach for evaluating relative powers of three selective texture feature extraction methods is carried out. In combining the results of the comparative study and the findings of other relevant studies it is found that the most accurate classification results can be obtained by incorporating texture features of Spatial Gray Level Co-Occurrence Matrix method with spectral features. However, if both classification accuracy and processing time are considered, incorporating texture features of the First Order Statistics method with spectral features appear to be the best option. There exists also a need to have a methodology to select a subset of features which is best suited for each particular classification problem. This study deals with this issue by modifying the Narendra-Fukunaga algorithm for selecting a subset from a full set of features in two-class feature selection problem. The modified algorithm ensures to select the globally best subset of features which might not be identified by the original Narendra-Fukunaga and saves the computing time significantly if compared with the exhaustive search method. This study also makes an attempt to modify the feature selection method for multiclass feature selection problem. A composite index which considers both the separability of closest class pair and the average separability over all classes is developed as a criterion for selecting feature subset. A technique to avoid problem of continuous increase of the separability values is also introduced. The effectiveness of the modified method for selecting feature subset is demonstrated through an experiment. Furthermore, the problem of high dimensionality when increasing number of features used would not only affect the accuracy achieved but also largely lengthen the classification time needed. Whereas, due to differences in underlying concepts or principles of available classification methods, it can be expected that changes of their performance along with the increasing number of features used would be different. Thus the decision to select which classification method should be used become more important. As an attempt to provide some tentative guidelines on this issue a comparative study is made for a number of most commonly used classification methods. The comparative study consists of experiments and analyzes supplemented by considering underlying conceptual principles of the studied classification methods. The research shows that the Maximum Likelihood Classifier (MLC) and the Binary Decision Tree (BDT) appear to be most consistent in providing highest classification accuracy at varying numbers of features used. Between these two methods however there exists some trade-offs. Hence, the MLC and the BDT methods are suggested as decision alternatives for analysts when selecting a classification technique. Based on the findings from the research a comprehensive framework is developed to systematically integrate spectral and texture features for classifying high-resolution images. The advantage of the developed methodology over the traditional ones is illustrated through an application carried out for classifying land use in the Nakhon Ratchasima Province of Thailand using the Thematic Mapper multispectral data. |