Research On Remote Sensing Image Classification Method Based On Structure And Deep Representation

With the rapid development of satellite sensor technology and the widespread application of remote sensing technology,it brings a lot of convenience and opportunities to everyone,but also fetches a series of new problems and challenges,such as: rapid increase in data volume,high redundancy;complex spectral and spatial structure;high data dimension.And the traditional feature extraction method has poor ability to characterize the complex space and spectral structure of the image.In this context,the framework of the dissertation will start from traditional feature extraction algorithms and deep learning network-based methods to study remote sensing image classification problems,and further improve the effect of remote sensing image classification.The main work of this paper is summarized as follows:(1)The joint sparse representation(JSR)algorithm has been widely used in the classification task of hyperspectral remote sensing image(HSI)in recent years.It can sparsely decompose the input pixels on an over-complete dictionary,and classify the test samples through the sparse representation residuals.However,the classification performance of the JSR classification algorithm is susceptible to the sparse representation dictionary’s ability to express samples.In order to solve this problem,this paper studies the hybrid metric representation from the three aspects of spectral band characteristics,spatial context information and texture structure characteristics,such as: information divergence,spectral angle,Bhattacharyya coefficient,etc.,to improve the JSR algorithm.And the extracted pixel similarity features are merged into the JSR algorithm through the decision function to further improve the separability of the feature categories,and then improve the accuracy of HSI classification task.The experimental results on the real data set of HSIs show that the three classification algorithms based on the JSR and hybrid metric representation can make full use of the spatial context information,texture structure information and spectral information of the image pixels.Compared with the original JSR algorithm,the classification performance is greatly improved.(2)Compared with traditional classification methods,convolutional neural network(CNN)has powerful automatic learning capabilities and can better explore deep semantic information in images.However,when the local receptive field of CNN is used to extract the features of the HSI,it may cause the inconsistency of the feature expression of the same pixel on the feature map,which will affect the accuracy of the classification result.Therefore,this paper will introduce the attention mechanism into the deep network model to improve the feature expression ability,and design a spatial-spectral attention aggregation network to solve the above problems.The network is divided into the spectral attention branch with squeeze-and-excitation module and the spatial attention branch with multi-scale selective kernel module.Among them,the spectral attention branch introduces an attention mechanism to assign different weight information to each spectral band to realize the selective learning of important features of the band;the spatial attention branch uses different size convolution kernels to extract the spatial features of the image,and the complementary spatial attention features of different convolutional layers are aggregated to avoid the problem of large differences in the amount of information expressed by different convolutional layers.The experimental results show that by introducing the attention mechanism into the spectral features and spatial structure information,the most representative spectral band features in the continuous spectrum dimension can be extracted,which greatly enhances the correlation between channels.At the same time,by extracting different levels of feature information in spatial dimensions,the spatial context information between adjacent pixels is greatly enriched.(3)Label information plays an important role in the supervised learning of high-resolution remote sensing(HRRS)image scene classification.However,due to the influence of many objective factors,the labels given in the data set may not be reliable and may contain noisy labels.Aiming at the problem of data set label uncertainty,this paper designs a noise label detection algorithm based on covariance representation.The main steps are as follows: First,using a pre-trained CNN model to extract the deep semantic features of the image scene,and then the dimensionality reduction method of principal component analysis(PCA)is applied to the output vector of the first fully connected(FC)layer to reduce the computational complexity.The noise training set is constructed to simulate the actual situation of label noise.Secondly,the covariance between the samples of the noise training set is calculated to form the covariance matrix.Next,the average value of the covariance matrix is calculated by row.And,the detection and removal of noisy labels is achieved by setting the decision threshold.Finally,the support vector machine(SVM)classifier is used to evaluate the improved training sample set to prove the effectiveness of the proposed noise label detection algorithm.Experimental results show that this method has made a great improvement in the detection of noise labels in the scene classification of HRRS images.In summary,for the problem of HSIs classification,this paper introduces a hybrid metric representation into the JSR algorithm and designs a spatial-spectral attention aggregation network to further improve the feature description ability of HSIs,and the superiority of the proposed algorithms are proved through experimental analysis.Aiming at the problem of HRRS image scene label noise sample problem,the noise label detection algorithm based on covariance matrix representation is designed in this paper,which greatly improves the quality of the data set and provides a new idea for the problem of scene classification data set purification.