Research On Hyperspectral Remote Sensing Imagery Classification Based On Multi-scale Integrated Classifier And Decision Fusion

With the rapid development of remote sensing technology,it has gradually become the indispensable data acquisition method in every application industry.Hyperspectral remote sensing is an extension after the multispectral remote sensing and one of the research focus in recent years.Hyperspectral remote sensing image can represent the spatial characteristics,at the same time,achieve the fine spectrum of the surface features,almost continuous spectral curve and huge amounts of data,many characteristics make it more potential applications.Many advantages of hyperspectral remote sensing image at the same time bring challenges of hyperspectral remote sensing image processing technology,such as huge amounts of data and multidimensional nature of hyperspectral remote sensing images to increase the complexity of data processing,also the phenomenon of “the same objects have different spectrum” and“the same spectrum have different objects” and “mixed pixels” easily lead to wrong points,etc.For multidimensional data features and pixel resolution problem,building high robustness and high precision classification model is an important technical problem to be solved in the field of high spectral remote sensing image classification.Space feature in image classification is important influence factors,by multi-scale spatial features enhance the classification results of robustness,coupled with the current several kinds of the most commonly used and classification of the classifier with better effect for image classification of different spatial scales,decision fusion with majority vote algorithm to choose the kind of get the most votes as the optimal classification results,in order to improve the classification precision.On the basis of this theory,this paper puts forward a method of hyperspectral remote sensing image classification based on multi-scale classifier integration decision fusion.The effectiveness and robustness of this method are tested by AVIRIS hyperspectral remote sensing image.This dissertation studies the performance of commonly used classifiers,spatial scale optimization,and decision fusion.Separately included the basic principles and advantages and disadvantages of SVM、SMLR、ELM、KELM.By weighted average value filtering of filtering window,the four kinds of classifier was observed respectively in three different training sample under the optimal window size,and obtain the optimal window scale corresponding to the overall classification accuracy of peak,explore space scale size links with the overall classification accuracy;Introduced the majority vote algorithm as decision fusion algorithm,the result of the multiple classifier classification of multi-scale window to vote the elements in the collection,to a class of with the most votes is regarded as the final classification results,namely "the minority is subordinate to the majority",the traversal of each pixel in turn yieldsthe classification result after the decision fusion,compare the differences between the decision fusion and conventional classification methods.According to the experimental results,the overall classification accuracy of this method is improved to some extent compared to other conventional methods under three different training samples.Therefore,this method has certain effectiveness and robustness.