Research On The Classification Of Hyperspectral Image Considered Both Efficiency And Accuracy

As the development of the remote sensing technology,we could be exposed to more information of earth surfaces.As the traditional application of remote sensing,the classification of remote sensing image has played an important role in precision agriculture,pollution regulation,disaster prevention and control,etc.,On another hand,as the development of the society,the application of remote sensing images classification is not only limited to the traditional fields,it has also been used to estimate the amount of insurance claim when major disaster occurred,the business field like the flow monitoring of customers in different scene,the energy field like the mineral resource planning,the high efficiency in collection of solar energy.The appearance of hyper spectral remote sensing technology brought the convenience and improvement in the process of obtaining sensing images which could provide us with the more elaborate spectral resolution so that we could have more abundant spatial position and texture information,we could detect the target and classify the ground features more easily as a consequence.However,the data acquired from a hyper spectral imager has the common problem of the big amount of data and the high dimension.Since the number of data is giant,the process to label the samples usually need a lot of time and is difficult to accomplish,also the introduction of huge samples could lead to the interference from irrelevant samples,which are so called noise samples.In addition,if we put all the spectral information into the process of training,according to the Hughes curve,could produce the redundant information and influence the accuracy.As a result,how to choose the optimal samples from the big data and used in the syncretic spectral-spatial data is the problem that needs to be solved in the practical application of hyper spectral sensing image classification.Based on the summary of above mentioned problems,this paper researched a method which selected the samples based on the self-spatial position information combined with the integrated spatial texture information.The experiments in this thesis are all carried out based on the platform of MATLAB using the hyper spectral data gathered from AVIRIS and ROSIS,the Indian pine and Pavia University sensing images which are very typical.The specific research is as follows:First,the paper started with the traditional methods of hyper spectral sensing image classification and proved the validity of the Least Square Twin Support Vector Machine(LSTSVM)by the classification consequence of Indian Pine data compared with others,which is the basic theory for our further research.Next,the paper divides the study into two parts,including the improvement of the calculation efficiency in classifier and the improvement of accuracy.In the model of 1-a-r classifier,the number of negative samples are much bigger than the positive ones,during the classification of every set of classifier,we keep the number of positive points fixed but select the negative points according to their contribute value to shrink the training samples and reduce the complexity in computing and the time consumed in training,also the process of manually marking the training samples is simplified.At the same time,correcting the result by the knowledge of neighboring samples,which tells us the samples in the same neighboring region used to have the same category.On the other hand,the paper processes the image data in the BEMD transform domain and extracts the principal components,then sends the components into the filter which is composed of a set of Gabor wavelets,integrates the output with the top three components acquired from the spectral information to form the new feature data to do the further classification experiment to help improve the accuracy of the consequence.At last,the two groups of means are combined to design a method which considered both the time efficiency and the accuracy is put forward,the experiments proved the effectiveness of the method proposed by this paper in the precision agriculture and city planning application.