Study On Feature Extraction And Semi-Supervised Classification For Hyperspectral Remote Sensing Imagery

With the development of imaging spectrometer,hyperspectral remote sensing image has been extensively applied in target detection,land use and land cover change,disaster monitoring,precision agriculture.How to improve hyperspectral image processing and intelligent analysis capabilities is the urgent issue.With the development of hyperspectral satellite and the popularity of UAV,the aerial remote sensing gradually into the public view,it is more convenient and get the rich hyperspectral remote sensing data,and the time period is greatly shortened,which provieds sufficient data support for hyperspectral remote sensing images.At present,most hyperspectral classification methods are based on spectral feature,spatial feature and the fusion of spatial and spectral feature together.When classifying on spectral feature,hyperspectral data are treated as an unordered distribution,and ignore the union of imagery and spectrum characteristics.For hyperspectral image,each pixel in hyperspectral image is an ordered set in a two-dimensional space,named geometric spatial feature,but not random distribution.With the development of hyperspectral remote sensing technology,the spatial resolution of hyperspectral image is increasing,and the shape and structure of the image is more and more abundant.Combine the spatial information with spectral information can improve the classification efficiency.At the meantime,feature fusion leads to dimension and classification expenses increase.Feature extraction is an effective way to solve the curse of dimensionality in hyperspectral classification.Most of the spectral feature extraction methods are mainly focus on low vision "pixel"level,and hard to human visual understanding.Morphological spatial features are always uncertainty influenced by parameters with high computational complexity.In order to classify hyperspectral remote sensing image efficiently with limited labeled samples,we deeply research on spectral and spatial feature extraction,and semi supervised classification algorithms.(1)We propose a novel spectral feature extraction method based on deep convolutional neural network.Put the original spectral data directly into the network,the pool layer and convolution layer connects alternatively to simulate of visual cortex complex.different layers of neurons are local connected.The frontier layer of neurons as the input of the posterior layer of neurons,we can obtain a number of features by applying several different filters.Then maximum pooling for each feature,and record the maximum value of each feature,all the features are spliced into a feature vector as the extracted high-level features.Similar to the abstraction of human vision,the extracted deep convolution features have higher differentiation and stronger expression ability.The advantages of spectral feature extraction based on convolution neural network are verified by our experiments.(2)We propose the spatial feature extraction based on multi-scale and multi-shape structure elements,which combines the features based on different structure elements,to avoid the single structure of EMP spatial information extracted from single structure element morphological operation.Conduct morphological transformation by different structure elements by Geodesic and reconstruction operations with step size through the structural elements of continuous scale increasing,to extract the image structure size and contrast.For different types of features,we use different structural elements to extract the most useful features,to provide data support for the subsequent classification.(3)We propose three semi-supervised classification algorithms,which are semi-supervised classifications with single feature-single classifier,single feature-multi classifiers,and multi feature-multi classifiers respectively.For the single feature-single classifier,we select high confidence samples into training set,modify the sample selection rules to increase more useful information;for single feature-multi classifiers,we utilize three different classifiers,select high confidence samples by the "minority is subordinate to the majority voting strategy",and add them into the training set after data editing.At the meantime,we combine with active learning theory to further improve the classification accuracy for semi supervised classification;for multi feature-multi classifiers,we only take into account for two classifiers.Train the two classifiers based on spectral and spatial characteristics,and pick the consistent samples predicted by two classifiers into training sets respectively until there are no consisted samples,then combine the spectral and spatial features to classify.Study on the semi-supervised classification algorithm for hyperspectral images,on one hand,make full use of useful information in the unlabeled samples can avoid Hughes phenomenon when labeled samples are limited.On the other hand,overcome some problems of semi-supervised classification algorithms for hyperspectral image,improve the classification performance under limited labeled samples situation.(4)Take the two typical hyperspectral remote sensing images,ROSIS Pavia University city and Salinas concise agricultural land cover as application case.It verifies that our proposed spectral and spatial feature extraction method and semi-supervised classification algorithms are efficient.