High Spectral And High Spatial Resolution Remote Sensing Image Classification Based On Deep Capsule Network

With the development of remote sensing detection technology,the obtained remote sensing data has high spectral and spatial resolution.Among remote sensing data processing techniques,hyperspectral image(HSI)classification and high spatial resolution remote sensing image(HSRI)scene classification attract extensive attention.Though HSI and HSRI have abundant spectral and spatial structure information,high dimensional data and complex spatial structure bring forward higher requests to feature extraction method.The thesis starts with complex feature distribution of the two categorical remote sensing data,develops robust classification models for HSI and HSRI based on deep capsule network.The model,which utilizes capsule’s capability of vector feature representation and feature integration capability of dynamic routing mechanism,can extract robust feature and further boost the classification performance.The main research contents in this paper are listed as follows:Firstly,the theory and classical model of deep learning especially convolutional neural network(CNN)are studied thoroughly.After that,the shortcomings of CNN are analyzed and deep capsule network is subsequently introduced.The characteristic of capsule network and dynamic routing algorithm are analyzed in detail.Based on the fully connected capsule layer,the paper proposes convolutional capsule layer by integrating the ideas of local connect and shared weight into capsule network layer,which has provided theoretical basis for building classification models based on capsule network for remote sensing data in the following parts.Secondly,HSI spectral and spatial-spectral classification based on convolutional capsule network are completed.One dimensional spectral capsule network classifier and three dimensional spatial-spectral capsule network classifier are designed based on one dimensional convolutional capsule layer and three dimensional convolutional capsule layer,respectively.The proposed model demonstrates capsule network’s effectiveness on HSI classification after the thorough analysis of model parameter.In addition,the proposed model achieves superior classification accuracy compared with other classical HSI classification methods on three HSI datasets.Finally,HSRI scene classification based on capsule network is studied.The paper transfer parameters of CNN which pre-trained on ImageNet dataset,the transferred model can transform the original pixel’s gray value to capsule’s vector representation and alleviate overfitting problem when the number of training sample is limited.Multiple layers’ features of CNN are extracted and convolutional capsule layer are used to further extract feature.The proposed method obtains more accurate classification results compared with other classical classification methods.