Research On Hyperspectral Image Classification Based On Semisupervised Collaboration

In recent years,with the rapid development of imaging spectrometer technology,hyperspectral remote sensing images have become a hot research area.Hyperspectral images contain hundreds of continuous bands,rich spectral information,and have been widely applied in many fields,such as mineral mining,environmental monitoring,as well as scientific agriculture.However,the classification of hyperspectral images has many difficulties due to the large number of bands and strong correlation between the bands.Applying fewer labeled samples to get better classification effect has drawn more attention among researchers.This paper summarizes the development of hyperspectral images processing at home and abroad in recent years,and also introduces some classical active learning algorithms,clustering algorithms and the collaborative active and semisupervised learning frameworks.We find that combining active learning and semisupervised learning can make full use of the available data and receive superior results.Introducing the wrong pseudolabels into the training procedure degrades the performance of the classifiers in the collaborative active and semisupervised learning framework.In order to solve this problem,this paper introduces the secondary screening schemas into one-fold collaborative active and semisupervised learning framework and double collaborative active and semisupervised learning framework,respectively.Therefore,we propose two novel frameworks.The first one is secondary screening algorithm and semisupervised collaborative framework(SSA-SCF).The second one is secondary screening algorithm and double verification semisupervised collaborative framework(SSA-DVSCF).These two frameworks can select representative and diverse unlabeled samples and avoid introducing wrong pseudolabels.We evaluate the experimental results by using the overall accuracy,average accuracy,Kappa coefficient and algorithm running time.From the experimental results,we find the proposed verification semisupervised frameworks cost too much running time.In order to solve this problem,this paper proposes a novel learning landscape features semisupervised collaborative framework(LLFSCF).This framework combines M-training algorithm and weighted spatial-spectral double layer SVM classifiers module(WSS-DSVM)and studies real-time landscape features.In this novel framework,we first apply SLIC(simple linear iterative clustering)based non-local superpixel segmentation algorithm to initially learn landscape spatial composition.Then,we apply WSS-DSVM module to obtain initial classification maps.To better characterize complex scenes of hyperspectral images,we quantizes both the landscape diversity and separability from the initial classification map,which achieves the availability of spatial details and the structural information of objects.Finally,we feed some patches with lower accuracy into M-training algorithm for further classification.In order to achieve an unbiased evaluation,we evaluate the performance of LLFSCF on two different scene hyperspectral data sets and compare it with that of three state-of-the-art hyperspectral image classification methods.The experimental results confirm the efficacy of the proposed framework.When secondary screening algorithms are used in our research,we find Entropy Query-by-Bagging(EQB)algorithm has the tendency to locally oversample in the complex areas.Therefore,EQB ignores relevant samples lying in uncertain areas which contains a small number of classes and results in slow convergence.This paper puts forward averaged normalized Entropy Query-by-Bagging algorithm(an EQB).SVM classifier and GBDT classifier are applied to verify the superiority of this algorithm,respectively.Then,we combine an EQB algorithm,MCLU algorithm and n EQB algorithm into a Multi-strategy fusion and multiple verification semisupervised framework(MF-MVSF).This framework mines more structural information from the unlabeled data selected by the different active learning algorithms and it can obtain effective and differentiated classifiers at the initial stage by using a few initial labeled samples.It has been proved that the proposed algorithm can effectively improve the classification results,reduce labeling cost,and eliminate the premature convergence caused by improper threshold setting.