| Hyperspectral remote sensing image has the characteristics of combining the spectral information determined by the material with the spatial information reflecting the shape and texture of the objects,so as to realizing the accurate classification,identification and attribute analysis of the objects,which greatly improve the cognitive ability to the objective world.Hyperspectral remote sensing technology has been widely used in geological mapping,vegetation survey,marine remote sensing,agricultural remote sensing,atmospheric research,environmental monitoring and many other fields.However,hyperspectral remote sensing data brings a lot of effective information,but also many challenges to hyperspectral image classification.The hyperspectral image classification mainly has the following problems: First,due to the complexity of the composition and distribution of the features and the impact of imaging conditions,hyperspectral images still exist the phenomenon of "different features with the same spectrum " and " the same objects have different spectra ".The second,the prevalence of mixed pixels on the hyperspectral image further increases the difficulty of classification of objects;Third,with the increasing of the image band,the demand for training samples in supervised classification method is also increasing dramatically,resulting in the pathological problems of small sample classification.In recent years,the development of the spectral-spatial classification,composite kernel classification,and semi-supervised classification algorithm and the spectral unmixing technology,provides possible solutions to the above hyperspectral image classification problems.In this paper,the hyperspectral image classification method of joint abundance information and spectral-spatial feature is studied from two aspects: supervised classification and semi-supervised classification.And two groups of aerial hyperspectral data with different spectral,spatial resolution and different ground complexity were used in the experiments to verify the validity of the proposed algorithms.The main research results and conclusions are as follows:1)This paper studies the extraction method of spectral-spatial feature based on morphology,and proposes the extraction process of extended multi-attribute profiles.Through the extraction of extended multi-attribute profiles,both the spectral and spatial features on the image are taken into account,which reduce the the impact of "different features with the same spectrum " and " the same objects have different spectra " on the classification.The spectral-spatial feature based on morphology could greatly improve the accuracy of hyperspectral image classification.2)In this paper,a class-based endmember extraction algorithm is proposed.The endmembers of spectral unmixing and classesof classification are often inconsistent,resulting in the problem that the end abundance information obtained by the unmixing can not be effectively applied to the classification.The algorithm first extracts the endmembers for the training samples of each category,endmembers corresponding to each category are obtained,and the endmembers of all categories are grouped together to form the endmember set,then the endmember sets are used to sparse Unmixing,and the abundance information corresponding to the category is obtained.The experimental results show that the proposed algorithm provides an effective solution for the problem of mixed pixels in the classification of objects,which can improve the accuracy of image classification.3)A spatial optimization algorithm based on Markov random field is studied.Considering that the distribution of objects in nature usually has a certain degree of spatial continuity,it is reflected in the remote sensing image,that is,the probability that the adjacent pixels are the same species type is the largest.The algorithm uses a Markov random field transcendental model to make the neighboring pixels belong to the same category,through the use of spatial neighborhood information to achieve the purpose of space optimization.The experimental results show that the spatial optimization algorithm can effectively remove the salt and pepper noise and improve the classification accuracy and effect.4)A composite kernel classification method combining abundance information and spectral-spatial feature is proposed.This method is more effective for image classification by fusing the characteristics of spectral-spatial information and abundance information at the kernel structure level.According to the composite kernel theory,the kernel of the abundance information and the kernel of the spectral-spatial information are combined to a composite kernel,and then the composite kernel is used to replace the original single kernel function.The experimental results show that the method of composite kernel can effectively integrate different feature spaces such as spectral,spatial and abundance,and obtain better classification performance than single kernel classification method of feature stacking.And can achieve the classification accuracy similar to the classification based on the feature selection,without the need of feature selection process,and improve the hyperspectral image classification efficiency.5)A semi-supervised classification method combining abundance information and spectral-spatial feature is proposed.Aiming at the small sample classification problem of hyperspectral remote sensing,this algorithm takes active learning by combining abundance information and posterior probability,and selects the most abundant samples from the perspective of category membership and mixed pixel for semi-supervised classification by improving the efficiency of classification.The experimental results show that the active learning method combining abundance information and posterior probability can achieve higher overall classification accuracy by using fewer training samples,which effectively reduce the workload of sample mark,reduce the time required for classifier training,and improve the efficiency of semi-supervised classification. |
PDF Full Text Request