Research On Classification Method Of Hyperspectral Remote Sensing Image Based On Capsule Neural Network

In recent years,the deep learning model has shown good performance in hyperspectral remote sensing image(HSI)classification field,especially convolutional neural networks(CNNs)because of its powerful feature extraction ability.Nevertheless,the classification methods based on CNN usually have complex network structures and require a large number of labeled samples to train the model.However,the number of labeled samples is limited which is a common bottleneck in the hyperspectral remote sensing image classification field.On the other hand,CNNs usually use max pooling operations to reduce the computation cost and advance the invariance of features which can capture more discriminative features but lose the relationship between the features of geographic objects.Thus,it can only learn shallow spatial features,while ignoring the important knowledge of spatial relationships and patterns for HSI.Moreover,owing to the complexity of HSIs,the scalar value used to represent features in CNNs shows poor representation ability.Considering the characteristics of hyperspectral remote sensing images,we introduce capsule neural network(Caps Net)to classify the hyperspectral remote sensing images in this paper.On the basis of making full use of the spatial spectral feature knowledge,this paper deeply excavates the inherent mechanism of the ground objects hidden in the hyperspectral image,namely spatial relationship and spatial pattern knowledge,and discusses the potential of Caps Net in the field of hyperspectral remote sensing classification.Compared with the state-of-the-art CNN-based methods,the classification accuracy and generalization ability of the model have been greatly improved.The main research contents and achievements of this paper include the following three aspects:(1)The non-local Caps Net(NLCaps Net)hyperspectral remote sensing image classification method that introduces the attention mechanism is implemented.Because the network of Caps Net is relatively shallow,it is difficult to obtain global information,especially for hyperspectral remote sensing image with complex surface environment,only extracting local information cannot achieve good classification results.In this paper,a novel NLCaps Net was proposed for HSI classification by combining the attention mechanism and Caps Net.The non-local block was used to obtain more information about the target and model the long-distance dependencies of hyperspectral remote sensing images in Caps Net,which uses global information of input images for hyperspectral remote sensing image classification.The proposed method can effectively enhance the classification accuracy with limited training samples.When 10% training samples are randomly selected from the Kennedy Space Center(KSC)?Pavia University(UP)and Salinas(SA)datasets,the classification accuracy of NLCaps Net is 98.86%,99.83% and 99.95% respectively.(2)A deep convolutional Caps Net(DC-Caps Net)hyperspectral remote sensing image classification method based on local dynamic routing is implemented.In view of too many parameters of capsule layer in Caps Net,we introduce the local connection and weight sharing to form a 3D convolutional capsule layer,and proposed a novel deep convolutional capsule network,which greatly reduces the amount of model parameters.Moreover,a lightweight decoding network based on deconvolution layer is proposed as regularization strategy,which greatly alleviates the risk of overfitting when the number of training samples is insufficient.When 3%,0.5% and 0.5% training samples are randomly selected from the KSC,UP and SA datasets,the classification accuracy of NLCaps Net is 95.97%,96.71% and 97.14% respectively.(3)A multi-scale deep feature aggregation Caps Net(MS-Caps Net)hyperspectral remote sensing image classification method based on feature fusion is implemented.Inspired by the success achieved by NLCaps Net and DC-Caps Net,we propose a capsule residual block based on residual connection.From the perspective of multi-scale feature aggregation,a deep local feature extraction module and a deep global are constructed based on the 3D convolutional capsule layer.The feature extraction module realizes the simultaneous extraction of spectral features,local spatial features,and global spatial features of the target ground object.In addition,the SE module is introduced into the shallow network of MS-Caps Net to refine the shallow features extracted by the model,which greatly improves the feature extraction and generalization capabilities of the model.When 3%,0.5% and 0.5% training samples are randomly selected from the KSC,UP and SA datasets,the classification accuracy of NLCaps Net is 97.67%,97.58% and97.84% respectively.