Active Learning with Support Vector Machines in Remotely Sensed Image Classification

Support vector machine (SVM) has been widely applied in the classification of remotely sensed image. How to reduce support vector number in SVM classifier so as to reduce classification time still an important open problem, especially in the case of mass data. To obtain fast classifier with high accuracy, an active learning schema is proposed in the SVM based image classification. Experimental results with synthetic data and multispectral remotely sensed images show that, compare with the SVM classifiers trained with whole training sample set in a time, the SVM classifiers obtained by active selection of training instances have much fewer support vector and can always achieve relatively higher accuracy.