Research On Hyperspectral Image Classification Method Based On Dense Convolution

As a three-dimensional image,hyperspectral image has hundreds of bands and contains a lot of information.With the development of spectrometer imaging technology and the improvement of remote sensing data processing technology,the current hyperspectral image processing technology can be applied to many aspects of production and life,such as biological threat monitoring,agricultural and forestry planting,atmospheric environmental research and marine research,military reconnaissance,and geological survey and many other aspects.The commonly used technology in these applications is the classification of pixel features of hyperspectral images,which is called hyperspectral image classification.It is a current hot research field and is of great significance to production and life.At present,in actual remote sensing image applications,due to the difficulty and high cost of labeling hyperspectral images,there are often insufficient labeled samples,which affects the classification accuracy.Therefore,how to effectively classify images under small-scale samples is the current research focus.Accordingly,this paper proposes a hyperspectral image classification algorithm based on dense convolution and conditional random field(Densenet-CRF)when the small-scale sample training set and test set are derived from a single image;In the case that the train set and the test set are derived from different images,a Hyperspectral Image Classification Based on Dense Convolution and Domain Adaptation(DCDA)is proposed.The main research contents are as follows:1.Convolutional neural networks used for hyperspectral image classification will have a gradient drop as the number of layers increases,which will make the classification results unsatisfactory.In addition,deep learning network training usually requires a large number of samples.However,due to the application of more cross-layer connections in the dense convolutional network,the problem of gradient descent can be well optimized,and the number of parameters can be greatly reduced,and the hyperspectral image classification algorithm based on dense convolution in the case of small-scale samples a better classification result can be obtained under the following,so a dense convolutional network is introduced for hyperspectral image classification.But at the same time,the current common hyperspectral image classification algorithms based on dense convolution generally do not make full use of spatial context information,but instead focus on local information,which will eventually lead to misclassification of some samples.Conditions follow the airport method can solve this problem well and make full use of global information to classify images.Therefore,this paper proposes a hyperspectral image classification algorithm based on dense convolution and conditional random fields(Densenet-CRF),and adds the Max Pooling layer as the output layer of the dense convolution network,so that each sample corresponds to a certain place the probability calculation of the object category is more accurate.First,the three-dimensional convolution kernel is used to synchronously extract the joint spectral spatial features and the layers are densely connected to obtain a dense convolution network,which reduces parameters and improves the accuracy of feature extraction.Then,the softmax layer is used to assign the categories of samples,and finally use the conditions the random field fully combines spatial global information to improve classification accuracy.And verify the effectiveness of the algorithm on the Indian Pines dataset and Pavia University dataset.2.In the actual remote sensing image application,due to the difficulty and high cost of hyperspectral image labeling,there are often no labeled samples in a hyperspectral image,or the number is very small,but another image with a similar feature category is high.Spectral images have sufficient conditions to label samples.However,there may be a spectral shift between two different images,and the conventional single-image classification algorithm cannot solve this problem well.Therefore,it is necessary to introduce a domain adaptive migration learning method when classifying cross-hyperspectral images.However,the structure of deep feature learning in the current common domain adaptive algorithm for cross-image classification of hyperspectral images will have the problem that the gradient decreases as the number of layers increases,which makes the classification result unsatisfactory.When the training set and the test set come from different images,in order to reduce the spectral offset,strengthen feature extraction,and improve the classification accuracy under small-scale samples,a hyperspectral image classification algorithm based on dense convolution and domain adaptation is proposed,and use it for accurate classification of hyperspectral images.First,build a dense convolutional network to learn the characteristics of the source domain.Secondly,the domain adaptive method is introduced to transfer the learned features in the source domain to the target domain to obtain the classification result.And verify the effectiveness of the algorithm on the Indian dataset and Pavia dataset.