Hyperspectral Remote Sensing Image Classification Based On CNN And GCN

With the development of sensor technology,the spectral resolution of remote sensing images is getting higher and higher.How to effectively classify hyperspectral remote sensing images has always been a difficult and hot topic in image processing.Traditional classification methods mainly use manual feature extraction,which is limited in feature expression capability and weak in generalization,and cannot meet the demand of high precision classification.Deep learning-based hyperspectral image classification methods can automatically learn and extract deep-level image features through a hierarchical perceptual network structure to achieve high-precision classification,thus gaining widespread attention from researchers.However,deep learning techniques still face many challenges in the field of hyperspectral image classification.For example,it is difficult for existing methods to fully extract spectral and spatial features in images,and they suffer from sparse label samples,numerous and redundant bands,and complex and poorly adaptable network models.Based on the above background,we focus on hyperspectral image classification methods based on spectral morphological features as well as deep learning,and have conducted experimental validation on several hyperspectral datasets to prove the effectiveness of these methods.The main innovation and research work of this thesis includes the following three parts:(1)A novel hyperspectral image classification method combining the Spectral Angle Mapping-Combination Characteristic Parameter(SAM-CCP)is proposed.This method combines the overall features of the reflection spectrum and the local absorption valley features.The spectral angle distance,which represents the overall feature of the spectrum,is first calculated,and then the Euclidean distance of the absorption valley feature parameters is calculated.The spectral angular distance is combined with the Euclidean distance,with adjustment of the latter’s open-square factor.This method solves the problem that the SAM cannot obtain the local spectral feature.By combining the global spectral feature with the local feature,the optimal spectral feature parameter combination can be selected automatically,and the classification accuracy can be improved.Compared with other spectral matching methods,the SAM-CCP method can effectively improve classification accuracy and has better applicability.(2)A novel hyperspectral image classification model based on the Enhanced Spectral-Spatial Residual Attention Network(ESSRAN)is constructed.This network combines spectral-spatial attention mechanism(SSAM),residual network(Res Net)and long-short term memory(LSTM)for deep feature extraction.The spectral-spatial attention mechanism can enhance the useful band and pixel information and suppress the useless band and pixel information.The use of Res Net can effectively prevent the gradient disappearing and exploding.The LSTM can learn the high-level semantic features in the spectral sequence.This method solves the problem that some networks,such as attention mechanism,ignore the semantic relationship features between bands,and solves the problem of few labeled samples by increasing the number of training samples using pixel clustering.Compared with other attention mechanism-based networks,this ESSRAN network can obtain higher classification accuracy with few samples,ensuring efficiency.(3)A novel hyperspectral image classification model based on Multiscale Feature Search-based Graph Convolutional Network(MFSGCN)is constructed.In this method,3D asymmetric decomposition convolution and 2D convolution are used to extract pixellevel features of hyperspectral images respectively.Then,the graph convolution network(GCN)is used to extract superpixel level features of different scales.The neural structure search method automatically assigns different weights to different scale features.Thus,more discriminative feature maps are obtained for classification.Compared with other GCN-based networks,the designed MFSGCN network can automatically select the layer with more feature information to extract features and achieve higher classification accuracy.