Research On Hyperspectral Image Classification Method Based On 3D Convolutional Capsule Networks

With the continuous development of non-contact detection,remote sensing image technology shows the trend of multi-spectral bands and high-resolution.Due to the fact that multi-dimensional structures contain abundant information,hyperspectral images have gradually become the core content of remote sensing technology.Target detection and classification are key components of hyperspectral image technology,which play important roles in environmental monitoring,agricultural management,military construction and urban planning.However,there are still some obstacles to classifying hyperspectral images,such as spectral information redundancy,complex spatial positional information,and insufficient labeled samples,which raise the difficulties of accurate classification.On the basis of deep learning theory,we propose models based on convolutional neural networks and capsule networks for classification,as well as exploit the jointly spectral-spatial information of the input data to extract features with higher discrimination,so as to increase the classification accuracy and overcome the limitation of labeled samples.The main research contents of this article include:Firstly,the specific process of classification is established through analyzing the structural information and evaluation criteria of classical hyperspectral data sets.As the research basis,the typical architectures and basic theories of convolutional neural networks are introduced in detail,and their advantages and disadvantages in the field of feature extraction are described.In addition,the theoretical researches of capsule networks are presented,as well as the principles and advantages of capsule neurons and capsule network structures are elaborated in detail.Secondly,on the basis of 3D convolutional neural networks,multi-region feature fusion method is utilized to build hyperspectral image classification model.Aiming at the problem of mixed pixels that may exist in the network input pixel blocks,the idea of multi-region branches network and feature fusion is employed to effectively extract features of hyperspectral images based on global spatial information and local spatial information,so as to improve the performance of the proposed network.Compared with the classical classification methods,our network has achieved better classification results in multiple data sets.Finally,according to deep capsule networks and convolutional neural networks,a new model is constructed which are composed of two multi-scale structures.Apart from improving the representations of features by utilizing vectors instead of scalar neurons,the width of the proposed networks is increased,and the capacity of ourmodels to distinguish various types of entities under the condition of limited samples is improved.Comparison results suggest that the proposed networks have a better capacity of classification.