Application Of Multiple-point Simulation And Rough Set Theory To Classification Of Remotely Sensed Imagery

Sampling is the fundamental of supervised classification for remotely sensed imagery, and the samples will directly affect the quality of remote sensing image classification results. Therefore, it is necessary to remote sensing image classification that the sample quality is evaluated, and thereby the uncertainty of the samples is depicted. Currently, most of the existing evaluation techniques are based on statistics, and in which the sample points are selected according to prior knowledge and experience. Under the same sampling pattern and the equivalent sample volume, the "real effect" of the sample data, which are used to train the classifier, can be validated and appraised until the images are classified. In other words, whether the spectral characteristics of data can be fully reflected by the samples, effectiveness of samples on classification would be known only after the images are classified. It is a meaningful works that how we can appraisal sample data to supervise and optimize the classification process before classification. Based on rough set theory, this paper proposed a method for appraising sample data quality by using "purity" of data. On The other hand, traditional methods for remote sensing image classification are generally based on the spectral information, and image's rich spatial structures and correlation information are failed to fully utilize. Although there have been some new ways to make use of spatial structure and correlation information for classification, such as context and texture, these methods have its own limitations. As a new inductive learning method, rough sets theory received widespread concern by its merits, such as "no priori assumptions on data are needed", "it can provide knowledge acquisition methods for incomplete and inconsistent data" and "easy understandability of knowledge obtained by it". In this paper, rough set theory is introduced into the sample appraisal.In recent years, the multiple-point geostatistics is gradually developed. Instead of varigram it uses training images to present the spatial structure characteristics of geological variables. Because it uses pixels as simulation unit the data are honored at their locations. So multiple-point geostatistics can be effectively used for modeling surface objects with complex geometric structure by sequential non-iterative algorithm with faster processing speed than object-base simulation algorithms.As the base of this research, remote sensing image classification process, multiple-points geostatistics and existing sample appraisal indices are introduced, a new sample appraisal method and a new method for remote sensing images classifying then proposed in this paper. The main research works are listed as follows.1. Quality evaluation of the samples based on rough set theory. First, some commonly used sample appraisal methods, especially Bhattacharya Distance and Transformed Divergence are summarized in this paper. Beyond the traditional statistical method, a new method based on rough set theory is then proposed in which the "purity" of the sample is calculated to evaluate the quality of samples. The experiment shows that there is certain linear relationship between this index and Bhattacharya Distance in the view of statistic.2. A new classification method for remotely sensed imagery: MLC+MPS. In order to introduce the structure information of data into classification, a new method based on multiple-point geostatistics and data fusion, MLC+MPS is proposed for remotely sensed imagery classification in this paper. The maximum likelihood method and multiple-point geostatistics are integrated respectively by maximum likelihood method and multiple-point geostatistics. Finally the two probability vectors are synthesized together by using data fusing methods and the classification result is generated. The experiment results show that the classification accuracy can be improved by MLC+MPS.