Research On Hyperspectral Image Classification Algorithm Based On Deep Learning

Hyperspectral images contain abundant spectral information and spatial information,hyperspectral remote sensing technology has been paid more and more attention in recent years.By reasonably using this information,accurate identification and classification of ground features can be achieved,so as to carry out quantitative and qualitative analysis.As the concept of deep learning becomes more and more perfect,a hyperspectral image classification learning algorithm based on deep convolutional neural networks has gradually become a research hotspot.Specifically,the extracted features often determine the quality of the classification results.By constructing a deep neural network suitable for the structural characteristics of hyperspectral images,more abstract and representative features can often be extracted and fine-tuned to Training deep neural network parameters to make them optimal,so as to achieve high accuracy classification goals.However,the problems of high dimensionality,large amount of data,feature redundancy,and correlation between spectra in the classification process have not been solved well.In this paper,from the aspects of data dimensionality reduction,feature extraction,band selection,and classifier optimization,three simple and efficient algorithms for improving the performance of hyperspectral image classification are proposed.Algorithm 1: Use local retention discriminant analysis to reduce the dimensionality of hyperspectral data.After processing,the hyperspectral data is filtered with a two-dimensional Gabor filter to generate spatial tunnel information,and the convolutional neural network is used to extract spectral features from the hyperspectral data to generate spectra.Tunnel information,fuse spatial tunnel information and spectral tunnel information,and input it into deep convolutional neural network.Taking full advantage of the advantages of local retention discriminant analysis and the high-performance computing characteristics of deep convolutional neural network algorithm,a hyperspectral image classification algorithm based on local retention dimensionality reduction convolutional neural network is proposed.Algorithm 2: Multi-scale convolution can detect the subtle changes between the spatial dimensions of the hyperspectral image and the local area pixels in the spectral dimension,and can be applied to the feature extraction of complex and diverse types of hyperspectral data.The three-dimensional neural network has the potential to capture local three-dimensional patterns,and is more suitable for the structural characteristics of hyperspectral images,which helps to improve the classification performance.The multi-scale 3D convolution model designed by combining the advantages of multi-scale convolution and 3D convolution extracts a three-dimensional image composed of pixels in a small spatial neighborhood(not the entire image)as input data along the spectral band to extract deep spectrum-Spatial features,train the deep classifier through Softmax loss,and predict the label category of each pixel.A multi-scale three-dimensional convolutional neural network hyperspectral image classification algorithm is proposed,which is helpful to solve the problems of low classification accuracy of hyperspectral images and easy confusion of similar spectral information.Algorithm 3: Using multi-index fusion dimension reduction to make full use of the rich spectral information and spatial information of hyperspectral images to reduce the loss of feature information.Combined with multi-scale dynamic convolution,it is extremely sensitive to the spatial changes of the hyperspectral image and the small changes between the pixels in the local area in the spectral dimension,which can be applied to the feature extraction of complex and diverse types of hyperspectral data.Finally,multi-label fusion conditions are used to construct the optimal band selection combination and make full use of multi-scale dynamic convolutional neural network to extract high-latitude depth features.Experimental results show that the proposed algorithm can simultaneously extract deep-level spectral-spatial features in hyperspectral images.At the same time,using the deep spectrum-spatial information in the hyperspectral image to improve the classification accuracy of the hyperspectral image,a hyper-spectral image classification algorithm with multi-index fusion dimensionality reduction and convolutional neural network was proposed.Performance is better than other algorithms.