Hyper-spectral Image Classification Based On Active Learning

The emergence of hyper-spectral remote sensing technology brings both new opportunities and challenges to traditional remote sensing technology. Due to the characterizes of high dimension, large amount of data and limited number of labeled samples of hyper-spectral remote sensing images, traditional remote sensing image processing method losing its utility. Aimed at the shortage of labeled samples problem in hyper-spectral remote sensing image classification, we introduce the active learning which belongs to machine learning in it. Active learning through active select unlabeled samples and marked them to expansion the original labeled sample set, and using the expanded labeled sample set to update the classifier, iterative this process until reached the stop conditions or the unlabeled sample set is empty, therefore to arrive a desirable classification accuracy under a labeled sample's insufficient situation.This paper combine support vector machines(SVM) and active learning and introduce them in hyper-spectral remote sensing image classification. Based on the existing classification model of active learning, through system studied basic theory of active learning and the characteristics of hyper-spectral remote sensing image, aimed at this two problems of using active learning in hyper-spectral remote sensing image classification, that is classification accuracy and sampling costs, proposed an appropriate classification model for each, in addition, analyzes the performance of the two combined models. The main contents and conclusion of this paper are as follows:(1) In the view of traditional hyper-spectral remote sensing image classification process, aimed at the simply sample selection criterion, which would easily cause informative samples lost, proposed a diversity sample selection criterion to measure sample information. Based on the diversity criterion, using minimum expected error criterion to select the desired minimum error samples, to maximum extent improve the accuracy of classification result. Verified by experiment with hyper-spectral remote sensing image, paper proposed algorithm can effectively improve classification accuracy, and can achieve a stable classification result under small sample situation.(2) As the large amount of data, especially unlabeled data, using traditional active learning method conduct hyper-spectral remote sensing image classification usually appeared a phenomenon of too much sampling cost, thus proposed a based on angle Locality-Sensed Hash(LSH) method for hyper-spectral remote sensing image classification. Instead of distance using angel to measure sample category, and mapping the calculated angel into LSH stack, which expected to maximum extent reduce the sampling cost. Through hyper-spectral remote sensing image classification experiment, the result shows that this proposes method could effectively reduce sampling cost and can also apply to the hyper-spectral remote sensing image classification with a large amount of data.