Hyperspectral Image Classification Based On Deep Learning

Hyperspectral imager is multi-band surface image captured by satellite or aircraft.It is widely used in many fields such as water quality testing,precision agriculture and geological surveying.In these applications,surface information is often very important,so we need to classify hyperspectral images.In recent years,with the rapid development of deep learning,neural networks have reached state of the art in many fields.In order to deal with the data in different fields,people successively invented a variety of models based on neural networks,such as convolution neural network,automatic encoder,long and short memory network.The flexibility of the neural network model allows researchers to adjust the structure of the neural network according to different data property in order to achieve the best effect.According to the property of hyperspectral image data,it is the focus of this dissertation to design a targeted neural network.This paper first analyzes the difficulty of hyperspectral image classification and verifies it with a simple experiment.To solve these difficulties as the goal,this paper draws on the idea of pre-training in the autoencoder and proposes a supervised pre-training method in hyperspectral images.In order to fuse spatial information and spectral information,this paper draws on the idea of parameter sharing in convolutional neural network,shares the parameters of multiple feature extraction networks,effectively reduces the dimension of input information and greatly reduces the number of network parameters.Combining the two ideas,this paper presents a neural network-based model to classify hyperspectral images,which can effectively solve the difficulties in hyperspectral classification,while leaving some flexibility.Experiments with traditional methods show that deep learning method is effective for hyperspectral classification problem.Experiments with other deep learning methods show that,compared with the direct use of existing neural network structure,our model are better.Finally,the article analyzes the model proposed in this paper in detail and verifies the necessity of pre-training and parameter sharing with experiments.It also shows the connection and difference between our method and convolution neural network and introduces the similarities between our model and decision fusion model.The main innovation of this paper is to design a deep learning model for hyperspectral image by using ideas of pre-training and parameter sharing.