Researches On Hyperspectral Images Classification By Multiple Features Convolutional Neural Network

Hyperspectral remote sensing images usually contain hundreds of spectral bands and achieve nano-scale spectral resolution,which make them contain rich spectral information of ground objects and have unique characteristic that other remote sensing images don't have.Therefore,Hyperspectral images(HSIs)have widely used in agricultural monitoring,geophysics exploration,military investigation and other fields.HSIs classification has always been an important research topic,and it is also a key technology to describe the detailed information of ground objects.In recent years,since deep neural networks have achieved great success in the field of image processing,more and more researchers have begun to use convolutional neural network(CNN)to classify HSIs.Typical CNN-based classification methods usually adopt image patches of HSIs as input to the network.However,fixed-size image patch in HSIs with complex spatial contexts may contain multiple ground objects of different classes,which will disturb the correct identification of the real category of the image patch by the CNN,and cause misclassification problem easily.In addition,traditional CNN-based classification methods only utilize the features of the last convolutional layer for classification without the full consideration of the information obtained by the previous convolutional layers.In view of the above two problems,this paper proposes two HSIs classification algorithms,and develops a software interface of the model test systems.The main contents of this paper are summarized as follows:1)Aiming at the problem that fixed-size image patch in HSIs may contain multiple ground objects of different classes,which will cause misclassification problem easily.This paper proposes a HSIs classification method based on multiple bias features convolutional neural network.According to the pixel values of different ground objects are different,its response in the network is also different,this method introduces a multiple biases module(MBM),which can decompose the feature maps of input patches into multiple bias maps(corresponding to different ground objects)by adopting multiple bias magnitudes.Then,the network flexibly select the feature maps by the self-optimization to achieve more accurate classification.Experiments show that by introducing the MBM,the misclassification problem can be effectively alleviated and the classification accuracy is improved.2)Aiming at the problem that the traditional CNN-based classification methods don't make full of multiple layer features.This paper proposes a HSIs classification method based on multiple features fusion convolutional neural network.This method includes two fusion strategies.The first is called multiple outputs joint decision fusion,which optimizes each convolutional layer in the network to create a side classification map.After that,the decision fusion is applied on these classification maps to obtain the final result.The second is called multiple features linear fusion,which linearly superimposes the hierarchical deep features of multiple convolutional layers,and then uses the fused feature to classify.The experiment results prove that the richly multiple layer features of CNN can be used to classify HSIs accurately.3)Aiming at the model test systems of the above two classification algorithms,this paper achieves software development and interface design by using the Pycaffe interface and Qt Designer tools in the environment of Python 3.6.