Research On Classification Algorithm Of Hyperspectral Objects Based On Convolutional Neural Network

Hyperspectral imagery is an image obtained by photoing a target over hundreds of spectral bands using a imaging spectrometer.The data of the hyperspectral image is in the form of a cube.Each band corresponds to a grayscale image,and hundreds of bands form a data cube.Therefore,for each pixel,there are hundreds of radiation values that are nearly continuously distributed from the ultraviolet to the near infrared(for examaple).Due to the above characteristics,hyperspectral images deeply reflect the physical characteristics of the target object.Therefore,hyperspectral images are widely used in the field of remote sensing,which can serve the military,agriculture,meteorology,and disaster warning.The main research of this thesis is the classification of hyperspectral images.Image recognition is a classic problem in computer vision.Its purpose is usually to determine the category or attribute of an object in an image.One of the most important tasks in hyperspectral image recognition is to determine the class of physical entities corresponding to each pixel,which we call hyperspectral image classification.From another point of view,hyperspectral image classification is also an image segmentation problem,that is,the whole image is semantically segmented by the spectral characteristics of the target,spatial neighbor relationships,texture structures,and so on.But in the field of remote sensing,we usually call it hyperspectral image classification.In the field of hyperspectral image classification,many methods have been proposed by the predecessors.From the decision tree model to the support vector machine model,they have achieved good results.However,when the deep learning method has achieved the effect beyond humanity in the image recognition task,more and more researchers have invested in its research.This paper attempts to apply the deep learning method to hyperspectral image classification to explore how to make deep learning achieve excellent results in hyperspectral image classification.This paper uses the most commonly used deep learning model in the classification model,namely the convolutional neural network to process hyperspectral images.Since there are many methods for classification in machine learning,the key problem of hyperspectral image classification is not the selection of classification models,but the high dimensional and high redundancy characteristics of hyperspectral data.The former leads to a huge amount of data in hyperspectral data,which brings challenges to the efficiency of classification;the latter leads to a large number of useless features that interfere with the final classification performance of the classifier.Therefore,we especially need to consider how to reduce the dimensions of hyperspectral images and how to extract effective features.Convolutional neural networks have the ability to acquire high-level semantic features and have a natural advantage in dealing with tasks with strong local feature correlation.However,in those classical network structures,deep neural networks are usually designed as end-to-end models.Feature extraction,classifier design,and classification decisions are often mixed together.In this case,it is not conducive to migrating the information learned by the model to other tasks.In order to solve these problems,we have made many improvements on the basis of the classical convolutional neural network,and proposed a novel method for the dimensionality reduction and classification efficiency.Both methods have achieved superior methods and have the potential to migrate to other tasks.Aiming at the band selection in dimensionality reduction,we propose a supervised band selection method based on one-dimensional convolutional neural network.The point is that in natural images,convolutional neural networks have the ability to preserve positional information of valid image features.Then in the hyperspectral image classification,we can select the band that effectively participates in the classification based on the positional information of the discriminative band recorded by the convolutional neural network.We replace the fully connected layer in the original network with the global average pooling to generate a band contribution map.With the help of contribution map generated in the classification process,we can estimate the contribution of each band to the classification process of the category.The contribution value decides the effective band to participate in the classification for each category.The band selected by our method has a good performance in classification accuracy on two commonly used hyperspectral databases.Not only that,because of its class sensitivity,the method can determine the band position sensitive according to the category.In other tasks,it may also have application value.Aiming at the problem of high band redundancy in hyperspectral images,we propose a convolutional neural network with simultaneous dimensionality reduction and classification.Our model simultaneously trains a one-dimensional convolutional neural network and a two-dimensional convolutional neural network.The data dimension reduction part and the one-dimensional convolutional network in the two-dimensional convolutional network share weights.One-dimensional neural network adopts only spectral information,which is used to extract spectral features of raw data.The input of the two-dimensional convolutional neural network is a hyperspectral image after dimensionality reduction by a one-dimensional convolutional neural network.Since the extracted spectral features and spatial features and classification are performed simultaneously,the spectral features extracted by the model maximize the spatial characteristics,and improve the convergence speed of the network and the efficiency of the model.The results on two commonly used hyperspectral databases show that the training speed of the model is 9.6 times higher than that of the three-dimensional convolutional neural network and 2.8 times higher than that of the two-dimensional convolutional neural network.The model has faster convergence speeds and competitive classification accuracy.