Research On Hyperspectral Image Classification Method Based On Sample Augment And Deep Learning

In recent years,with the prosperous development of the aviation industry and the continuous innovation of hyperspectral imaging technology,the analysis and processing of hyperspectral remote sensing image has attracted more and more attention.As one of the most important components in the field of hyperspectral remote sensing technology,the task of hyperspectral image classification has aroused a heated discussion.In the task of hyperspectral image classification,feature learning is usually the focus and difficulty of the research.The framework based on deep learning can highly integrate and abstract the low-level features,which has an outstanding performance in the field of computer vision and attracts the attention of many researchers in the field of hyperspectral imagery.However,the contradiction between the data characteristics of hyperspectral image and the prerequisite for the application of deep learning-based method also poses challenges for further research.In order to better use the deep learning-based model for feature learning and classification,this paper starts from the sample augment,addressing the hyperspectral labeled sample scarcity issue.Meanwhile,for the sake of learning the multi-scale semantic information and improving the classification accuracy of hyperspectral image,this paper improves the existing hyperspectral image classification methods based on convolutional neural network.The main work of this paper is summarized as follows: 1.We propose a hyperspectral image classification method based on superpixel sample augment and Generative Adversarial Networks,in which the superpixel sample augment expands the dataset of hyperspectral image,and the improved Generative Adversarial Networks further learns the features of the expanded samples.This method makes use of the advantages of superpixel segmentation technology in local spatial information extraction,and the multi-scale spatial information of hyperspectral image is also fully learned while the samples are expanded.On the premise of truly limited number of labeled samples,this method highly integrates the spectral and spatial information of hyperspectral images and improves the classification effect.2.We propose a hyperspectral image classification method based on adaptive dilated convolutional neural network.Based on the traditional convolutional neural network,an end-to-end double branch spectral-spatial feature learning and classification framework is constructed.The spectral branch extracts the spectral information of the sample efficiently and adequately by 1×1 convolution.By replacing the convolution layer and pooling layer of the traditional convolutional neural network by the designed adaptive dilated convolution layer,the spatial branch learns the multi-scale contextual information of the hyperspectral image on the premise of maintaining the spatial resolution and receptive field.Combining with the structure characteristics of hyperspectral image,the proposed method pays attention to the convolutional content in the convolution,thus the obtained deep spectral-spatial feature has a stronger representation.Compared with the classification method based on convolutional neural network,the classification performance is obviously improved.3.We propose a hyperspectral image classification method based on selective-kernel convolutional neural network.The proposed method constructs an end-to-end spectral-spatial feature learning and classification framework by simply stacking the selective kernel modules.It can adaptively adjust the size of the receptive field of neurons,improve the adaptability of the network to the targets of different scales in the image,and better capture the multi-scale contextual information.The method can effectively learn the spectral-spatial characteristics of hyperspectral image,and the design concept is simple,which can be easily applied to different data sets.Compared with some existing classification methods based on convolutional neural network and the method proposed in the previous work,the classification accuracy is improved obviously.