Application Of Deep Learning In Hyperspectral Remote Sensing Image Processing

In recent years,with the rapid development of remote sensing technology,hyperspectral remote sensing become a hot issue of research due to their characteristics of multiple bands,high resolution and large amount of information.Relevant academic researchers have paid more attention on on hyperspectral image dimensionality reduction and classification technology research as well.Traditional hyperspectral image processing methods can only extract the shallow features of the image,but ignore the deep features of the image.Meanwhile,this kind of method cannot take into account the spatial and spectral features of the hyperspectral image at the same time,so it is difficult to get a better classification effect.Aiming at the above problems,this paper uese deep learning method to process hyperspectral images.Firstly,hyperspectral images has the relatively strong correlation between bands,which lead to high redundancy.In order to solve this problem,the depth subspace clustering method is used to select the band of hyperspectral images to improve the efficiency and accuracy of subsequent processing.The method uses convolutional autoencoder to extract the potential characteristic representation of the band.A self-expression layer is introduced between convolutional encoder and decoder,and the feature extracted from the encoder is used to measure the similarity between bands to obtain the similarity matrix.Finally,the spectral clustering analysis of the similarity matrix is carried out to obtain the final band selection results.In order to improve the perfprmance of the depth subspace clustering network,we introduce the activation function Leaky Re LU to prevent gradient disappearance caused by the inactivation of large areas of neurons;the method of small convolution kernel superposition is used to extract deep features and fit the data better;short skip connect is used to fuse the shallow feature and the deep feature across the layer to accurately and quickly extract the potential feature of the band.The experimental results on the public data sets show that,compared with other traditional methods or models,the selected bands of the method based on the depth subspace clustering are more representative,which verifies the effectiveness of the method.Secondly,in order to effectively utilize the abundant feature information provided by hyperspectral images,the method based on convolutional neural network is used to classify hyperspectral remote sensing images.This method takes 3d image as input and uses thespatial information and spectral information of hyperspectral image to classify it and improve the classification accuracy.To improve the performance of the network,introduce Inception structures to enhance network scaling adaptability;the residual block of bottleneck structure is introduced to prevent network degradation and gradient disappearance;batch normalizationis adopted to prevent gradient disappearance and accelerate network convergence;ELU activation function was used to avoid neuronal necrosis;The global average pooling layer is used instead of the full connection layer to reduce parameters and prevent the occurrence of overfitting.The experimental results on the public data sets show that,the classification accuracy of the method based on the convolutional neural network in this paper is superior to that of the traditional convolutional neural network,neural network and support vector machine.Finally,hyperion hyperspectral images in the east of Shijiazhuang were processed by the above hyperspectral processing method based on deep learning.First of all,hyperion image is pre-processed by repairing ingstripe,removing broken lines,removing invalid bands and correcting atmospheric.Next,the band selection of the pre-processed hyperspectral images was carried out,and combined with geochemical datas,the soil heavy metal pollution in this area was evaluated.The experimental results show that this method has high accuracy in evaluating soil heavy metal pollution.