Research On Multispectral Remote Sensing Image Classification With Pattern Recognition Methods

With the development of sensor and platform technology of Remote Sensing, the data of Remote Sensing have a huge raise in terms of quality, quantity and variety. Remote sensing has increasingly been used as a source of information for characterizing land use and land cover change at local, regional, and global scales. The government makes decisions based on the thematic information extracted from the remote sensing images. While, inherent heterogeneity, complexity, inaccuracy characteristics of geospatial data and fuzzy expression of geographic phenomenon present challenges for extracting thematic information.Pattern recognition is the most significant field of artificial intelligence, which is combined with many theory and technology such as statistics, cognitive science and information science. In the field of remote sensing, pattern recognition methods based on computer technology for land use/land cover classification of multispectral remote sensing images are the most commonly used means for thematic information extraction, especially when dealing with the practical application.The thesis takes the practical application of pattern recognition methods in remote sensing image processing and other areas as the clue, and tries to find the linking point between the pattern recognition methods and thematic classification of multispectral remote sensing images. Based on the research and development of multispectral remote sensing image classification, starting with spectral characteristics of multispectral remote sensing image classification and improving the accuracy of remote sensing image classification for the target, the thesis proposes a new feature extraction method for multispectral remote sensing imagery using local histogram, and a novel segmentation method for multispectral remote sensing imagery using the K-means classifiers ensemble coupled with local combination histogram. Finally, the thesis applies object-based classification method to classify the segmentation results. The main achievements and conclusions involve the following aspects,1. There are an intensive studies and comprehensive summaries on the existing pattern recognition method for extracting the thematic information from remote sensing images.Firstly, the thesis briefly describes the the basic workflow of pattern recognition system. Secondly, the thesis deeply describes two steps, feature extraction and classifier design, including PCA (Principal Component Analysis), ICA (Independent component analysis), supervised classification, unsupervised classification, multi-classifier aggregation and other pattern recognition methods for remote sensing image classification. The two steps are the working emphases of the thesis. Finally, the thesis briefly describes the method for evaluation of remote sensing image classification.2. The thesis proposes a method for feature extraction from multispectal remote sensing images based on local histogram.The histogram of remote sensing image is a statistic variable of spectral information in nature. The existing image segmentation methods based on histogram is a kind of multi-threshold operation on histogram. It can not adapt to the images with complex contents due to the poor separability of the whole image histogram. Furthermore, a significant but usually ignored problem is that the land area represented by a pixel comes from the surrounding terrain. So the local histogram is used as the segmentation features. The local histograms of every pixel are computed in every band image which is quantized. Then the local combination histogram (LCH) of every pixel is obtained by concatenating the local histograms of the pixel in different bands. The Quantizing process could reduce the redundant information and also keep the borders’information for segmentation. The local histogram contains both spectral information of pixels and neighborhood spatial information implicitly. Experimental results followed show that the proposed feature extraction method could achieve more accurate segmentation maps for multispectral remote sensing image segmentations.3. The thesis proposes a novel segmentation method for multispectral remote sensing imagery using the K-means clustering ensemble coupled with local combination histogram (LCH).This thesis takes the successful application of multi-classifier ensemble in multispectral remote sensing image classification and presents a new multispectral remote sensing image segmentation method based on K-means clusting ensemble. Firstly, all bands will be partitioned into nonoverlapped subsets of bands, which are viewed as nearly independent information sources. Then bands conmbinations are generated from these subsets. Secondly, LCHs which describes pixels are computed in every band combination. The identical K-means algorithm is applied to the LCH features in different band combinations. Then several relatively coarse segmentation maps are obtained. Finally, a final K-means procedure is performed to get a refined segmentation reslult. Since different band combination highlights different types of land cover, the aggregation of the segmentation results in different band combinations could account for more land cover classes and get higher recognition accuracy. Synthetic and real Landsat-7 ETM+ satellite images are used to illustrate the value of the proposed mehtod in the practical application. Experiment results suggest that the proposed method could restrain oversegmentation and undersegmentation to a great extent and exceeded other methods. 4. The thesis implements the object-based classification method by performing supervised classification algorithm to the image objects (patches) in the segmentation results.The conventional per-pixel classification methods always produce Salt-and-pepper phenomenon which is caused by lacking in spatial information, texture information, structure information and contextual information. It militates against thematic mapping seriously. Object-based classication, however, can overcome these disadvantages. Synthetic and real Landsat-7 ETM+ satellite images are used in the classification experiments. Experiment results suggest that the object-based classification base on our segmentation method could improve the performance of maximum likelihood classifier (MLC) to a great extent. The fact that the better segmentation results cause the higher classification accuracy makes object-based classification an assessment index of segmentation. So the classification experiments results also indicate that the segmentation method proposed in previous section exceeded other methods quantitatively.The main contributions of the thesis are summarized as follows. The proposed LCHs for multispectral remote sensing image serve as the feature which takes into account both spectral information and spatial information. The LCHs overcome disadvantages of the traditional feature extracting methods such as PCA and ICA which are very sensitive to noise and outlier. A segmentation method for multispectral remote sensing image based on strategy of K-means clusting (unsupervised classification) ensemble is proposed. Since the recognition accuracy of land-cover types is improved, oversegmentation and undersegmentation are restrained significantly. The object-based classification method is implemented based on the image objects from the segmentation results and achieves a higher classification accuracy.