Random Forests And Its Application To The Classification Of Remote Sensing Image

With the development of remote sensing technology, how to classify theremote sensing image automatically、accurately and rapidly is the hotspot.Deriving of mangrove remote sensing image’s training sample is very hard, soit’s a big challenging to the accuracy of remote sensing. On basis of randomforests, together with TM features, an improved Random Forests(RF) wasproposed which is more automation.RF was widely used in classification of remote sensing image for itsstability and suitable for small samples. In order to improve the accuracy andefficiency of classification of remote sensing image, several works was doneon the basis of random forests:First, in order to distinguish classes which are more similar, a completerandomness feature combination random forest algorithm(CRFC_RF)wasproposed, which can exploit combination information of TM imageautomatically. On the basis of feature linear combination, CRFC_RF addedrandom features selection and random feature space selection, which haslower generalization error and higher classification accuracy.Second, for the purpose of improving classification efficiency of RF,Clonal Selection(CS)was introduced to optimize RF, after optimization, RFbecomes smaller and faster,together with higher classification accuracy.At last, incorporating characteristic of RF, an unlabelled sample selectionmechanism was proposed based on the margin maximization. As the resultshown, the selected unlabelled samples have active contribution to RF undermargin maximization selection mechanism.For the purpose of verifying the effectiveness, all algorithm proposed wasvalidated on UCI data sets, and the classification result of remote sensing alsocompared to traditional algorithm.