Research On Land Cover Remote Sensing Image Classification With Pattern Recognition Methods

The development of remote sensing technology trends to have three highlighted features which are high spatial resolution,high sepectral resolution and high time phase resolution with three direction of multi-platform,multi-angle and multisensory.People can get bigger and bigger remote sensing image data.At present remote sensing technology are widely used in territorial recourses investigation,crop yield evaluation,urban planning,land cover change,traffic monitoring,ecological environment monitoring,military investigation and seismic surveillance.Land cover change and land utilization gives government important evidence when making decision on land.It is also related to human life and production.Production of large scale of thematic map and macroscopic detection of land cover change play very critical role in ecological environment monitoring and social economy.Accurate real time extraction of remote sensing date classification is not only related to date quality and resolution but also related to principle of classification method.Pattern Recognition is a new branch of science back to 1960^th.It is interdisciplinary field which has close connection to computer science,statistics,cognitive science and cybernetics.Development of Pattern Recognition and computer automatic control technology spread the remote sensing image with computer intelligence interpretation.The ultimate aim of the remote sensing image with computer intelligence interpretation is to recognize and classify different earth surface.This thesis focuses on applying Pattern Recognition in remote sensing image classification.Beginning with remote sensing image features the thesis tries to find out combination point between Pattern Recognition and remote sensing image classification in order to promote accuracy of remote sensing image classification.It takes ALOS/PALSAR,PSM remote sensing image as source of experimental data and use remote sensing image space to classify earth surface from remote sensing image.Firstly,the thesis will introduce pattern recognition theory in detail within thematic classification.Several improved methods are to put forwarded to make SVM kernel function better after overall research of support vector machine,SVM with classification experiment and result analysis of remote sensing image.Then it provides a new remote sensing image sorting algorithms which combines sample distance measure methods and pattern recognition such as SVM,KNN and ELM.After carrying out classification experiment and comparing sorting result with other pattern recognition,the thesis suggests partition method of remote sensing image based on joint of fuzzy clustering and SVM to fulfill object-oriented remote sensing image classification.Achievements of the research are followings:1.General principles of pattern recognition are discussed and summarized in remote sensing image classification field.At first,principles of pattern recognition are introduced in detail in the thesis.Second,common statistic pattern recognition such as supervised classification,unsupervised classification,half supervised classification and multiple classifiers are fully described to emphasize application of statistic pattern recognition in the field of remote sensing image classification.At last,the thesis states evaluation method of classification accuracy.2.Two kinds of improved methods of optimizing SVM Kernel function.The thesis makes a comprehensive study on principles of SVM classification method and Kernel function,especially on the selection of SVM Kernel function and optimal parameter.It also provides two complex Kernel functions to improve SVM Kernel function.Optimizing Kernel function can achieve two goals.On the one hand it complements each other's advantages;on the other hand sample measure function gives consideration to differences of sample luminance and angle.Feasibility effectiveness is shown by experimental data from remote sensing image classification experiment and contrastive analysis of traditional Kernel function's classification performance.3.The thesis puts forward two sample distance measure methods and ELM-SVM multiple classifiers model based on traditional classification algorithm and mixture decision rule,referencing successful application of multiple classifiers in remote sensing image classification.This multiple classifiers method which was purposed in the paper effectively overcomes sluggishness of KNN and problems of optimal k parameter selection,at the same time effectively improve the classification accuracy of ELM method.All kinds of index of remote sensing image experiments show that these algorithms we give in the thesis are better than other classification algorithms.4.Through study of object-oriented remote sensing image classification technology,this research comes up with pixel automatic segmentation based on FCM,and combined with SVM realizes object-oriented remote sensing image classification.The difficulty of object-oriented remote sensing image classification technology lies in segmentation of remote sensing image.At present there are several partitioning algorithms.Some are histogram;some are regional;some are edge detection;some are map.This thesis gives a simple and practicable pixel automatic segmentation classification combined with FCM and SVM.Robustness of our method is proved by large scale experimental data from contrastive study of other method effectiveness.