Research On Hyperspectral Image Classification Based On Active Learning

Nowdays,hyperspectral remote sensing has gradually become the forefront of remote sensing development.Hyperspectral images contain hundreds of narrow and continuous wavebands,providing rich and delicate spectral information for identifying and classifying features and it have been widely used in mineral exploration,environmental monitoring,resource exploration,and agriculture.However,hyperspectral image have large amount of data,large number of bands and strong correlation between bands,which pose a huge challenge to the classification of hyperspectral images.In the absence of labeled samples,how to get better classification performance with less labor cost becomes the key issue of hyperspectral image processing,which also attracts more and more researchers' attention.This paper summarizes the research results of scholars at home and abroad in recent years,using the active learning algorithm based on entropy bagging to build a sample pool,and then assist in the unsupervised clustering process of secondary screening pool division,marking samples have more more information and more representation.Experiments on hyperspectral images show that a multi-level screening strategy can obtain similar classification results with less labor cost.Aiming at the decline of the false label verification effect and the convergence of the threshold in the case of the joint active learning and semi-supervised learning label verification framework,an active learning framework of multi-verification is proposed.Using the samples collected by different active learning strategies,a differentiated calibration classifier is constructed,and the pseudo-label is verified by multiple checks to increase the confidence level of the pseudo-label so as to reduce the misjudgment the loss of information samples in the experiment to be proven.The proposed algorithm can effectively improve the convergence rate and reduce labor costs,and put an end to the premature convergence caused by improper threshold setting.