Hyperspectral Image Classification Based On Improved Spatial-spectral Features And Convolutional Neural Network

Recently,with the advancement of science and technology,great progress has been achieved in the field of Earth remote sensing.It is essential of precise classification of features for the rational use of the Earth’s resources.Compared with ordinary images,Hyperspectral images（HSIs）not only contain spatial information,but also have more abundant band information.These spatial and spectral features provide data support for accurately classify the ground objects.However,there are still some problems,such as the huge amount of spectral information sacrificing the spatial resolution to some extent,and the strong correlations between spectras causing redundancy and so on.In view of the above difficulties,the main problems to be addressed and works are as follows:（1）Aiming at the problems of insufficient spatial features and the utilization of spatial consistency at the data-wise,a method of Data-wise sp Atial regio Nal Consistency Enhancement（DANCE）is proposed.Firstly,HSI is divided into many blocks.For each image block,undirected graph is generated by spectral graph theory,and the image is decomposed once by spectral graph wavelet transform（SGWT）to obtain image components in different frequency domains.Secondly,Gaussian filter is performed on the obtained image components to remove Gaussian noise in different frequency domains.Then,through the inverse SGWT,all the filtered components are reconstructed to obtain the image with spatial consistency enhancement.Finally,the optimized Diverse Region-based Convolutional Neural Network（DRCNN）is used for classification.Experimental results on several public datas show that the proposed method enhances the spatial consistency,removes noise on multiple scales,and effectively improves the classification effect of DRCNN and other methods.（2）There are strong correlations between spectral bands of HSIs,which lead to the spectral feature redundancy and difficult classification.Aming these problems,a band selection method based on spectral correlation is proposed.Firstly,the correlation coefficient matrix C between spectral bands of HSI is calculated,then the value representing self-correlation in the matrix(i.e.,the value of C_ii)is set to 0.Secondly,the mean m_cof matrix C is calculated.Thirdly,the threshold u（0（27）u（27）1）is setted.Finally,the number is counted if the correlation coefficient is greater than m_c.The band is deleted if the ratio of number is greater than u.All bands are opreated as described above until the proportion v（%）of the band to be retained is reached.This method preserves the effective spectral information to the greatest extent and reduces the spectral redundancy at the same time.The experimental results show that the proposed method improves the classification accuracy by increasing the difference between spectras,saves the calculation cost and reduce the the of CNN training.（3）Based on the method of feature re-extraction for misclassification regions,a classification model of ASRCNN（Auto-Selection Region CNN）is proposed.Firstly,each of the HSI block is divided into five regions.And the regions are pre-classified by the first CNN branch to obtain the OA（Overall Accuracy）.Secondly,area with the smallest OA is selected as the target region,and the window of size c×c is traversed in this region to calculate the misclassification degree _eP of each window.Then,the window of largest _eP is selected as the re-Selection Region（SR）and will be sent to the second CNN branch.Finally,the results of the six regions are fused to obtain the classification results.This method achieves the task of making the network pay more attention to the pixels that are not accurately classified,and reducing the redundancy of repeated window extraction in DRCNN.The experimental results show that the improved ASRCNN achieves good classification results.