Research On Deep Learning Models For Hyperspectral Image Classification

With the improvement of spectral resolution of remote sensing imaging sensors,then it is possible to identify and interpret the characteristics of landscape objects by means of the spectral attributes,that not only promotes the development of remote sensing technology,but also makes hyperspectral remote sensing one of the state-of-the-art research fields.Regarding hyperspectral image has the characteristics of multiple spectral bands,the higher spectral resolution,and the rich spectral information,it has the better ability of recognition and extraction of land cover information,along with some disadvantages such as the low spatial resolution,the information redundancy,and the curse of dimensionality.Additionally,hyperspectral image classification techniques have rapidly evolved with the fast development of machine learning,pattern recognition,and artificial intelligence.That is,how to intelligently take advantage of the finer spectral signatures included in the hyperspectral data,and an effective solution to this problem is to meet the requirements of specific applications whilst taking the classification performance,the generalization capability,and the intelligent level into account.Artificial intelligence techniques realize the intelligent interpretation of land cover on the earth's surface,and the automatic remote sensing image processing with the help of deep learning methods which are the cutting-edge studies at present.First,deep learning stems from the researches of artificial neural network;second,the remotely-sensed data are analyzed and interpreted by multi-layer perceptron.Furthermore,deep learning has two obvious characteristics:(1)automatic feature engineering,and(2)multi-layer abstraction,which are helpful to improve the final classification accuracy.As far as deep neural networks and feature learning concerned,on the one hand,deep learning technique adopts multi-layer perception to extract the features from low-level to high-level gradually,which is helpful to improve classification accuracy;on the other hand,as for different applications,the high-level feature representation can be automatically learned from the massive data through feature engineering,and the implicit information can be expressed effectively.Deep learning models can achieve the remarkable performance in the task of hyperspectral image classification.In particular,convolutional neural networks can mine the land cover information in spatial domain,and capture the spatial relationship of features between pixels.Therefore,deep learning models have the great advantages in feature extraction.Based on this,hyperspectral image classification with deep learning methods using the real hyperspectral data set is presented in this study.The main contributions of this study are stated as follows:(1)The research on capsule network in the field of hyperspectral image classification.Through choosing hyperspectral image as data object,the model performance of capsule network in the task of hyperspectral image classification was explored,while focusing on the performance of hyperspectral image classification and the capability of feature extraction of capsule network.(2)The research on deep residual network and densely-connected network in the field of hyperspectral image classification.Three residual learning models were designed to study the performance of hyperspectral image classification with respect to different models.(3)The research on neural architecture search in the field of hyperspectral image classification.Based on the automatic construction and search of deep learning models,this study explored the paradigm of architecture,network design,model training and hyperparameter optimization,as well as the selection of models.(4)To design and develop a deep learning framework for hyperspectral image classification task.Correspondingly,taking deep learning theory and technology as the focus and hyperspectral image classification as the motivation,then making full use of the ability of deep neural networks in information extraction and feature representation to label hypespectral data samples.That is,taking the deep learning models to conduct hyperspectral feature mining and data classification,and obtaining some useful findings and merits,which mainly include:(1)Capsule neural network with limited training samples to conduct hyperspectral image classification.This study proposes a novel capsule network that adapts to less training samples,so as to design and build a deep network framework,and then compares it with a typical convolutional neural network.The experimental results indicate that the proposed capsule network can achieve relatively high classification accuracy with a small amount of training samples,and the higher confidence for the maximum prediction probability.(2)Block-wise structurized residual network to conduct hyperspectral image classification.This study uses the comparable side-by-side paradigm of the framework design to evaluate different deep learning models,and finally overcomes the problem that the deeper convolutional neural network may lead to lower precision.At the same time,considering the features in the previous layer could optimize the number of features generated per layer and prevent information redundancy,which is important for deep learning models to better adapt to hyperspectral data.The experimental results demonstrate when trained by limited training samples with the same parameter total in terms of the ability of feature extraction of the deep convolutional neural networks,the densely connected network takes less training time than the residual residual network having no improvement of classification accuracy.(3)Weight-shared fast neural architecture search to conduct hyperspectral image classification.This study designs a fast neural architecture search network for the improvement of performance and accuracy of hyperspectral image classification to promote the level of automatization and intellectualization for deep feature learning in hyperspectral image classification task.The experimental results show that the neural architecture search network proposed in this study could make full use of the sample information in comparison with the artificially designed optimal deep convolutional neural network,even if only a small number of training samples can be obtained in terms of over accuracy,average accuracy,and Kappa index,and apparently more scientific and effective in the design,construction,search,and generation of convolutional neural network models.(4)The design and development of light-weight abstract framework for hyperspectral image classification based on deep learning techniques.The goal of Hyperspectral Networks(HSN)proposed in this study is to develop a Tensor Flow slim-like framework that enables the construction,training,and evaluation of deep convolutional neural networks regarding the application of hyperspectral data classification.That is an advanced abstraction framework or library that enables a higher level of intelligence in terms of algorithm integration,innovation discovery,and technical iteration.Through the implementation and integration of a variety of deep convolutional neural networks,our experiments indicate that the abstract framework takes the advanced features of deep learning techniques and the unique characteristics of hyperspectral image data into account.In this study,the application of deep learning techniques in the task of hyperspectral image classification uses the real hyperspectral data sets from three outdoor fields that reflect the state-of-the-art technological level.To realize the application of deep convolutional neural networks in the hyperspectral image classification task,and achieve accurate and fine land cover by designing and developing a high-level abstract framework for specific applications.That is,the accurate mapping,iterative update and improvement of the developed prototype of the hyperspectral image classification framework HSN.The experimental results demonstrate that the improved convolutional neural network architectures and the trained classification models can significantly improve the object recognition ability in hyperspectral image classification based on the studies of three state-of-the-art deep learning models.The derived classification accuracies have a good prospect,which further promote the development of hyperspectral remote sensing for the interpretation of land surface.