Hyperspectral Image Classification And Band Selection

Hyperspectral image has been attached great significance to due to its rich visual information.Because of the wide application which ranges from the space detection to marine inspection,hyperspectral image(HSI)influences our life on daily basis.However,with the development of life technology,people become unsatisfied with the current image resources.Therefore,the available HSI gradually develops from the perspective of space,spectral and time dimension.Among these,the research on spectra is still the most popular.With the augmentation of number of spectra,compared with traditional image,HSI can better reflect the natural characteristic of image.However,this also lead to the increase of image information,which means the image processing has been a vital problem with such easy access to HSI.Based on this,researchers are assigned with higher demands on the exploration of HSI.To satisfy the need of HSI processing,people on one hand tackle this problem by hardware updating to achieve real-time hyperspectral image processing;on the other hand reduce the HSI dimension and integrate the HSI information by software intervention.Further,this intervention especially pay much attention to hyperspectral image classification and band selection which can lay a strong foundation for the subsequent HSI processing.HSI image classification is a basic work in remote sensing.Among different problems,this application is always with high application value.The problems ranges from object tracking,anomaly detection,land utilization,and marine resource inspection.Among these works,HSI image classification in most time is viewed as an the preprocessing of the other works.Among most hyperspectral image processing,the researchers need first classify the whole HSI globally,and seperate different type of interest reasonably before making good use of them.From this perspective,the classification work is of close relation with the subsequent applications;If the classification result is unsatisfying,the later work will meet with unnecessary obstacles.Also because of this,it is not hard for us to find the significance of HSI classification in the remote sensing area.On the other hand,with the development of ROSIS,the imaging process can detect hundreds of thousands of HSI bands which can cover a wide range of spectra.Based on these data,the HSI can be viewed as ‘image cube'.Faced with such cube,reducing the spectral information has became a hot topic and band selection is just an effective dimension reduction method.This paper mainly concentrate on HSI classification and band selection,and commit the carry on the research from different views by proposing robust and effective analysis algorithms.The main innovation point is as follows:1.The current HSI classification method in most time is based on the consideration of spectral information regardless of the spatial information.Based on this problem,the first innovation work is to propose a stepwise iteration conditional method(ICM)for HSI classification,which can achieve better classification result.To be more specific,this method re-define the data term and smoothness term in ICM method,which improve the energy function in Markov Random Field(MRF).By separating the data term from the smoothness term and optimizing them one by one,the weights of these to terms are balanced.We also propose a SIFT-like feature,which consider the spatial relation during the feature extraction process.An effective spatial restriction is also introduced to improve the classification accuracy.This work is mainly verified on Indian Pines image.With the proposed classifer,the accuracy reaches 96.08%.2.The current band selection method mainly consider the bands individually which fail to consider the relation among bands.Based on this problem,the second innovation work is to maximize the correlation among bands to achieve high efficiency and low loss during the band selection process.To be more specific,this work first propose the dual-spectral angle feature based on the characteristic of HSI.By highly extracting and expressing the HSI,the needed context information of HSI bands are obtained.The intrarelation among bands are also explored and dual clustering framework is proposed.The immune colony model is utilized to obtain the final objective bands.The effectiveness of this method is mainly verified on Salinas scene.The SVM method is used for the later classification of HSI and with the selected bands the classification accuracy remains at90.02%,which is almost the same with the accuracy of HSI classification when all bands are utilized.3.The current feature selection method for HSI in most time cannot solve the curse of dimension problem gracefully and reducing the dimension of HSI without information loss cannot be achieved.Based on this problem,the third innovation work is to propose a multi-task sparse representation model.This model breakthrough the restriction in the traditional band selection which cannot acquire the global information of HSI.To be more specific,this work construct the band clustering framework to cluster initial HSI bands.By introducing the band representation model,the method represent the bands in forms of tasks.By sparsely re-constructing the initial HSI bands by selected bands,the effective dimension reduction is achieved.By erasing the redundant information with MRF,HSI is classified efficiently.This work is mainly verified on Pavia university image,with SVM classifier,the classification accuracy remains at 92.52%,which is almost the same with the accuracy of HSI classification when all bands are utilized.