The Research And Application Of Remote Sensing Image Segmentation By Fuzzy Clustering

By processing and analyzing the remote-sensing image, remote-sensing image segmentation is the technology and process to extract the target feature. The fuzzy clustering algorithm used in remote sensing image segmentation becomes one of research hotspots in recent years. The2006year LandsatETM+RS image about Meicheng Forest-farm in Dongzhi county of Anhui Province was used as the main data sources. In the support of RS and GIS, we use Erdas9.0to process layer stack, geometric correction and image cropping to preprocess the remote sensing images. The improved fuzzy clustering algorithms are applied to remote sensing image segmentation. The experimental results show that the proposed new fuzzy clustering algorithms produce the good performance. According to the actual needs of the project, the remote-image segmentation algorithms have been implemented. And the demonstration system of remote-sensing image segmentation has been designed and implemented.In sum, the main contributions of this dissertation are given as follows:Firstly, two kernel-based fuzzy clustering are presented:kernel fuzzy cluster centers separation (KFCCS) and kernel-based Inter-cluster separation (KICS) clustering. KFCCS is an improvement of fuzzy cluster centers separation (FCCS) algorithm. KFCCS can map input data points to a high-dimensional feature space where clustering unlabeled data is carried out. By using kernel method KFCCS can deal with linear non-separable problem better than FCM and FCCS. Experiments show the better performance of KFCCS. At the same time, KICS is an improvement of inter-cluster separation (ICS) algorithm. KICS can map input data points to a high-dimensional feature space where clustering unlabeled data is carried out. By using kernel method KICS can deal with linear non-separable problem better than FCM and ICS. Experiments show the better performance of KICS.Secondly, two fuzzy clustering algorithm based on optimized parameters are presented:an improved possibilistic clustering algorithm with optimized parameters (IPCAOP) and a generalized noise clustering independent of parameters (GNCIP). IPCAOP is an improvement of the combination of possibilistic clustering algorithm (PCA) and possibilistic c-means clustering (PCM). Since PCA and PCM are very sensitive to good initialization and they have undesirable tendency to produce coincident clusters, an improved possibilistic c-means (IPCM) algorithm was proposed to solve the problems of PCM. However, the performance of IPCM depends heavily on the parameters. IPCM must compute the parameters twice, so it is time-consuming. To deal with the problems of PCA and IPCM, and apply to remote sensing image. IPCAOP is proposed to combine PCA and IPCM. Our experimental results show that the proposed algorithm compares favorably with fuzzy c-means clustering (FCM) and IPCM from the aspect of Remote Sensing Image Segmentation. GNCIP is an improvement of the combination of generalized noise clustering (GNC) and PCA. GNCIP is proposed to deal with the shortcoming of GNC algorithm depend heavily on parameters and FCM must be performed until termination to calculate the parameters for GNC algorithm. With a nonparametric method GNCIP calculates the parameters in GNC objective function. So GNCIP algorithm does not depend on the parameters that GNC holds and clusters data faster than GNC algorithm. Experiments and simulation on two man-made data sets and two real data sets show GNCIP deals with noisy data well, cluster centers are closer to real ones, clustering accuracy is improved and clustering time is reduced.Thirdly, remote sensing image about Meicheng Forest-farm is segmented in four algorithms mentioned above. Then the efficient evaluation of the segmentation result is posed. At the end, a kind of algorithm with segmentation highest efficiency is choosed to segment image.Fourthly, a system of remote-sensing image segmentation has been implemented. As to the actual requirements of the project, the improved algorithms are implemented in system. By using object-oriented programming ideas and the vxWidgets of cross-platform graphical interface library, a demonstration system of remote-sensing image segmentation has been designed and developed.