Hyperspectral Images Classification Based On Deep Learning With Small Training Samples

Hyperspectral image(HSI)containing both spatial and spectral information,which brings a significant opportunity for land-cover classification and also poses a challenge for classification algorithm.Recent years,deep learning based HSI classification approaches have achieved significant breakthroughs.In deep learning based HSI classification method,a large number of labelled samples are needed for training and just very limited labelled samples are available in HSI,which has been a bottleneck to further improve the classification performance.This dissertation aims to solve these issues by building and optimizing deep network models to finish HSI classification task under the small sample condition,and verify the effectiveness of our proposed methods on several sets of hyperspectral datasets.The main contributions of this dissertation are as follows:(1)Considering that the existing methods of extracting spatial-spectral features from HSI by deep learning model neglect spectral features to a certain extent when introducing spatial features,a dual-channel network structure combining one-dimensional residual network channel and two-dimensional residual network channel is proposed.The combined dualchannel network structure automatically extracts the spatial and spectral features of HSI by taking advantage of autonomous learning of deep learning.Secondly,a semi-supervised co-training framework and sample selection mechanism based on spectral and spatial features are proposed to overcome the shortage of manual labeling data in HSI.In each iteration of the co-training process,the two-channel network is trained separately from the spatial-spectral characteristics,and some unlabeled samples are selected for each other for the next iteration.Finally,the advantages of this method are verified by testing on three published hyperspectral datasets and selected Hyperion datasets.(2)A semi-supervised deep learning framework based on three-dimensional lightweight network and deep clustering is proposed to solve the problem that the network scale is too large,the parameters are too large,and a large number of labeled samples are needed when threedimensional convolutional neural network is used for HSI classification.Firstly,threedimensional lightweight network is used to extract deep features and classify them.Then,an approximate rank clustering algorithm is applied to the deep feature clustering to generate a large number of pseudo labels of unlabeled samples.Finally,the lightweight network is finetuned by minimizing the dual-loss through the use of true labels and pseudo-labels.Clustering model and classification model promote each other,and they are updated iteratively in the form of collaborative learning.The experimental results show that the training of deep convolution network using pseudo label samples is effective.In the case of fewer training samples,experiments on three challenging hyperspectral data sets show that the performance of this algorithm is significantly better than the most advanced deep learning-based algorithm and traditional hyperspectral classification algorithm.(3)To overcome the over-fitting problem brought by the increasing depth of network in 3D convolutional neural network under small samples.An improved dense convolution network is proposed,which uses multi-scale dilated convolution.It is not a traditional scaling operation to learn features of different scales.The dense connection can connect the threedimensional feature maps learned from different network layers,so that multiple scale features can be obtained without increasing the network depth,and the network parameters can be reduced while over-fitting is effectively suppressed.Secondly,when existing depth models reduce the dimension of hyperspectral data,they usually need to reduce the dimension of the original HSI in the spectral space first,and then combine the spectral information with the spatial information extracted after dimension reduction to obtain the spatial characteristics.In view of the complexity of the process,the amount of calculation is large,and the loss of certain spectral information,which affects the accuracy,a novel spectral attention mechanism is proposed,which can selectively emphasize the spectral features with large information and suppress the small amount of spectral information.Finally,the experimental results on three hyperspectral datasets show that the proposed algorithm combining the three-dimensional multi-scale dense convolution network and the spectral attention mechanism is superior to the most advanced hyperspectral classification algorithm.(4)To overcome the problem that the mixed pixels affect the accuracy of hyperspectral classification caused by low spatial resolution,a novel hyperspectral classification framework is proposed,which introduces spectral unmixing into multi-scale dense networks as a supplement to the classification results.Considering the fact that mixed pixels often result in the mixing of easily separable samples and difficult samples,a network design of multiple intermediate classifiers with early-exiting mechanism is proposed.This design can shorten the test time without losing the classification accuracy,which brings considerable benefits to computational requirements and final performance.Secondly,aiming at the contradiction between training parameters of three-dimensional convolutional network and less training samples,a 3D/2D convolution based on the spectral-spatial features is proposed,which can make the network contain fewer 3D convolutions.At the same time,more spectral information can be obtained by using two-dimensional convolution to enhance feature learning,thus reducing the training complexity.Finally,the advantages of the method are verified by the four published hyperspectral data sets.Experimental results show that this method has better performance than other methods deep-learning based methods and traditional hyperspectral classification methods.