Study On Remote Sensing Classification Of Land Cover Based On Multi - Classifier Integration

Remote Sensing Image classification is one of the key techniques ofinformation extraction, is the remote sensing data conversion for the geographyinformation and the knowledge technical core.And the high accuracy remote sensingimage classification technology si a goal which the remote sensing applicationdomain unceasingly pursues for a long time, so how to improve the classificationaccuracy is the key problem in the application of remote sensing image.Multiple calssifiers integration thought and integration method research as animportant part of pattern recognition,has been widely applied to signalprocessing,fingerprint reader,and many other areas, and gradually rise in remotesensing image processing. It provides a way for more accurate classification resultsthat using the diversity of the existing classifiers by combining multiple existingclassifiers.The study first analyzed the remote sensing iamge classification researchdomestic and foreign progress,especially the classification of multiple classifiersintegration technology.On this basis, the experimental area of Landsat use ofShangrila County.(1) Sub classifier selection.First of all, the remote sensing images ingeometriccorrection, fusion, crop were used in pre-prcessing the image;Secondly,bythe selsected ROI to supervised classification of6seed classifier,and finallyselsedted3kind of sorters(Maximum Likelihood classification,Neural Networkclassification,Mahalanobis Distance classification) according to the sub-sorterclassified result and the performance to participate in the integration.(2) Sub claddifier integration.development of combination rules to integrate thevoting method,the maximum probability method,using ENVI/IDL to realizemultiple classifier for remote sensing image classification.The experimental resultsshow that:the overall accuracy and Kappa coefficient of multiple classifiersintegration than a single classifier classification accuracy highest MLC respectivelyto85.13%and0.82. (3) Integrated classifier impact assessment.Analysis integration effect andperformance,the results show that:on the basis of the accuracy of singleclassifier,integrates classifier performance correlation negative correlation betweenclassifiers.Experiment results shows that the multiple classifiers ensemble compares thetraditional classification method to be able to increase the classified precision,andhas good scalability,and by setting different combination rules can improve theclassification performance of integrated classifier.However,the data source of multi classifier integration process is directly to theresults of sub-classifiers treatment, therefore,for training of the classifiers in thebuild process optimization of sample selection,integration,a classifier is an optionand a multiple classifiers integration method and combination reules still need tocarry out in-depth research.