Hyperspectral Image Classification Based On Extended Morphology And Active Learning

With the development of hyperspectral remote sensing imaging,hyperspectral remote sensing processing has gradually become a research hotspot in remote sensing,and hyperspectral image classification is one of the most important fields in hyperspectral remote sensing processing.Compared with general remote sensing images,the hyperspectral images contain more information which can be used for classification,which meanwhile causes information overload.When there are not enough training samples to train the classifier,the classification accuracy tends to increase and then decrease with the increase of hyperspectral data dimension.The high dimensionality of hyperspectral image and the difficulty of marking training samples pose great challenges to classification.Therefore,how to improve the classification accuracy with fewer training samples by combining spatial and spectral information is one of the research hotspots.In order to solve the problem of using spatial information effectively and the lack of training samples in hyperspectral image classification,this thesis presents several classification methods which are based on extended morphology and active learning.The main work is as follows:(1)A technology of hyperspectral image classification based on extended morphology with single shape structural elements and active learning is proposed.In this method,spatial features are extracted by extended morphology with single-shape and multi-size structural elements,which are fused with spectral features to make full use of spatial and spectral information.At the same time,active learning is applied to reduce the need for training samples.In the active learning,the sampling strategy combining AP clustering with MCLU criterion ensures that the most useful training samples for classification will be selected.Compared with other technologies,this method can obtain better classification results.(2)A hyperspectral classification method based on extended morphology with multi-shape structural elements and active learning is proposed.In order to fully explore the spatial information,the method selects structural elements with multiple shapes and sizes to extract the extended morphological features of hyperspectral images.At the same time,the classification method of active learning is adopted to reduce the demand for labeled samples.The sampling strategy combining image partitioning and MCLU criterion is proposed for active learning to select samples with both uncertainty and diversity.The experimental results show that this method can achieve better classification effect with fewer labeled samples.(3)A new method of hyperspectral image classification based on genetic algorithm is proposed.Structural elements in extended morphology are usually selected according to experience or a large number of experiments,which is time cosuming and can not abtain the optimal combination in general.Considering optimizing the most suitable combination of structural elements for each hyperspectral image using genetic algorithm,two hyperspectral image classification methods are proposed.One is a hyperspectral image classification method based on genetic algorithm to optimize the combination of morphological structural elements.The method directly utilizes the spectral feature and the extended morphology with the combination of structural elements optimized by genetic algorithm,and randomly selects some samples for classification.The other is the hyperspectral image classification method based on genetic algorithm optimization of morphological structural elements and active learning,which combines the morphological characteristics of structural elements optimized by genetic algorithm with active learning to achieve a better classification effect with fewer training samples.Compared with the other techniques,it is proved that the proposed methods can make full use of spatial and spectral information and achieve better classification results with fewer training samples.