Study On Reliable Classification Methods Based On Remtely Sensed Image

Remotely sensed image classification is an important technique for extractingthematic mapping information. Due to the complexity of interactions in the naturalenvironment as well as the natural environment and remote sensing spectroscopy,surface information extracted from the sensor spectral signal includes uncertainties.Many uncertainties are inevitably introduced and spread during the process ofclassification. The nature of uncertainty must be known when using these dataincluding uncertainty, reducing its impact on the classification accuracy and obtainingreliable classification methods to improve the classification accuracy. The objective ofstudy is to provide reliable remotely sensed image classification methods. The processof remotely sensed image classification being as the starting point, uncertaintiesimpacting on the classification accuracy are studied in the main parts of theclassification processing, and reliable classification methods are presented to reducethe uncertainties in the results so as to improve the final classification accuracy. Thestudy has important theoretical significance and application value in extractinginformation from the remotely sensed image, remote sensing applications and spatialdata uncertainty. The body mainly includes effect of training samples on remotelysensed image classification accuracy, reliable remotely sensed image classificationmethods and reliable sampling methods for accuracy assessment, etc. The study willprovide a new method for improving the remotely sensed image classificationaccuracy. The details are as follows:(1) Relationships between quantity, quality and sampling method of trainingsamples and classification accuracies were concluded from experiments are as follows:1) Different classification methods’ responses to different size of samples areinequable, as well as the same method’s response to the same size of samples, thefinal classification accuracies trendings show a degree of volatilities, but the mean ofaccuracies is relatively stable when the size of samples is above certain size.2) Thequality of samples had major Effect on the classification accuracy, the accuracies aredifferent using different classification methods based on different quality of samplesunder different quality criterions. Different classification methods’ responses todifferent quality of sample under different quality criterions are roughly the same.Different classification methods’ responses to the same quality of samples under thesame quality criterions are different.3) Remotely sensed image classificationaccuracies are inevitable different based on training samples using different sampling methods, but the mean of classification accuracies can describe the true classificationresult based on training samples using different sampling methods. Stratified samplingmethod is better the non-stratified sampling method, and point sampling method isbetter than cluster sampling method.(2) Considering the mixed pixels being as the major factor resulting in the lowclassification accuracy and most of mixed pixels distribute on the boundary ofdifferent classes, the dissertation introduces fuzzy topology theory and presents anovel fuzzy-topology integrated support vector machine (SVM)(FTSVM)classification method for remotely sensed images based on the standard SVM. First,the optimal inter-correlation coefficient threshold value is applied to decompose animage class in spectral space into the three parts: interior, boundary, and exterior infuzzy-topology space. The interior class pixels are then classified as predefinedclasses based on maximum likelihood. The exterior-class pixels are ignored. Thefuzzy-boundary-class pixels which contain misclassified pixels are reclassified basedon the fuzzy-topology connectivity theory. As a result, misclassified pixel problems,to a certain extent, are solved, and the classification accuracy is improved.(3) The dissertation presents a new eigen-values-based multiple classifierscombination, in which, considering the classification capacity for every pixels basedon different classifiers, first, a probability vector matrix for every pixel is obtainedbased on the probability vector output from each classifier, then, the differentclassifiers’ weights on a pixel are decided based on the characteristics of theprobability vector matrix, a bigger weight will be attached on the good performanceclassifier, as a result, the classification accuracy and stability are improved.(4) The dissertation presents many remotely sensed image classification methodsby fusion of multiple spatial features. The details are as follows:1) Presenting a remotely sensed image classification method based on MarkovRandom Field-based Fuzzy C-means Clustering Algorithm (MFCM), the proposedmethod combines the results of a pixel wise spectral classification and a segmentationmap, aiming to improve classification accuracy, when compared to pixel wiseclassification only, in which the neighborhood information of pixel are fullyconsidered, first, the homogeneous regions are extracted using MFCM from theremotely sensed image, thereby, the spatial contextual information of pixels can beobtained from the homogeneous regions, then, the final classification result map isobtained by combining homogeneous regions map and the results of a pixel wise spectral classification and a segmentation map by majority voting.2) Presenting a remotely sensed image classification method based on SpatialAttraction-based Markov Random Field (SAMRF), pixels’impacts on the center pixelare inversely proportion to their distances from the center pixel under the Tbler’s FirstLaw of Geography, and spatial attraction (SA) model is introduced to express thenon-linear relationship at a distance about the interaction among the pixels, theEqual-weighted MRF (EWMRF) is improved to accord with the fact.3) Presenting many remotely sensed image classification methods by fusionspectral, textures and pixel shape features. The remotely sensed image is classified bydifferent fusions of Gabor wavelet, Gaussian Markov Random Fields (GMRF) andGrey-Level Co-occurrence Matrix (GLCM) textures and proposed PAI, LW, Solidityand Extent pixel shape features to improve the classification accuracy. Experimentalresults indicate that classification accuracy can be greatly improved by fusions oftextures, pixel shape indices or textures and pixel shape features, and improvementsof every class in the image are different by different fusions of textures or pixel shapefeatures. Generally, GLCM performs better than the other textures. Alone pixel shapefeature cannot improve every class’s classification accuracy, thus, multiple pixelshape features should be fused for remotely sensed image classification to obtainsatisfactory classification accuracy. Feature Selection always is performed in terms ofclassification by fusion of multiple features.(5) The dissertation presents Spatially Balanced Sampling (SBS) andCluster-based spatial stratified sampling for obtaining uniform and representativetesting samples, thus, the reliability of accuracy assessment result is improved.