Hyperspectral Data Classification Based On Extreme Learning Machine Theory

The characteristics of hyperspectral remote sensing data in high-dimensionality,mass, heterogeneous and with small amount of labeled samples bring a big challengeto the research on hyperspectral image. Extreme Learning Machine (ELM) is a newlycame up machine learning algorithm, which the input weights and bias of hidden layerof the network are random assigned, the output weights of the network can be gotclosed form, so it shows fast speed, high classification accuracy, and goodgeneralization performance in large-scale data classification.Based on the study ofELM theory, this paper aims at issues of hyperspectral remote sensing data such ashigh-dimensionality, limitted labeled samples, affected by the random noise and so on,and proposes sparse ELM, semi-supervised ELM, and Bayesian ELM learning model,in order to achieve fast and accurate classification of hyperspectral data.The studymainly includes the following aspects:Firstly, aims at dimensionality reduction and learning for hyperspectral data, thispaper proposes a sparse ELM model and corresponding learning algorithm. Sinceperformance of learning machine is closely related to data characteristics, we combinefeature selection of hyperspectral data and the choice of learning machine model,which comes down to sparse representation problem, and propose sparse ELM modelsand learning algorithm, which finish the data dimensionality reduction and structuraloptimization simultaneously. Sparse structure can not only reduce the computationalcomplexity of the network, and may lead to better generalization ability. We alsoextend the model to the deep structure，and construct a sparse deep ELM model andthe corresponding learning algorithm, in order to achieve fast and accurate learning forhyperspectral data. It is tested on hyperspectral data. The results show that this methodcan achieve fast and accurate classification on low-dimensional data.Secondly，aims at the phenomenon of mixed pixel and the high cost of labeledsamples in hyperspectral data, this paper proposes semi-supervised ELM algorithmbased on refined clustering assumption. It modifies clustering assumption commonlyused in semi-supervised classification, makes use of the information contained in theunlabeled samples to establish manifold regularization under the refined clusteringassumption, and proposes a kind of hyperspectral data classification technology basedon semi-supervised ELM and refined clustering assumption. It can achieve accurateclassification when very few labeled samples are used. Several experiments are done on hyperspectral data such as Indiana, and the results show that this method is not onlysuperior to other methods on the classification accuracy under the same condition, butalso time complexity is low.Thirdly, considering the mixed pixels and noises existing in hyperspectral data,this paper proposes a spatial-spectral Bayesian ELM model and the respondingalgorithm. Due to the impact of various systematic and random noises in the process ofhyperspectral imaging, we model to noise, construct a more robust Bayesian ELMmethod by introducing Bayesian techniques. It bases on the spatial consistencyassumption, jointly uses hyperspectral spatial-spectral information, constructs a spatial-spectral Bayesian ELM model and the responding algorithm. It corrects the outputweights of the network through Bayesian estimation, thereby enhances theclassification accuracy. The effectiveness of our proposed method is evaluated viaexperiments onAVIRIS data, and the results prove its excellent performance.The research was supported by the National Basic Research Program of China(973Program) under Grant No.2013CB329402, National Nature Science Foundationof China under Grant No.61072108,60601029,60971112,61173090, New CenturyExcellent Talents project No.NCET-10-0668, The ministry of education doctoralprogram funds (20120203110005) and Higher school subject innovation engineeringplan (111plan), No. B0704.