Research On Hyperspectral Image Classification Method Based On Deep Learning And Spatial Spectrum Features

These recent years have witnessed the rapid development of imaging spectrometer,which enables hyperspectral remote sensing images to provide researchers with a large amount of information about the feature category.At the same time,various processing technologies of hyperspectral remote sensing images have also been advanced,and the classification technology of hyperspectral remote sensing images has also made new breakthroughs under this background.The classification technology is one of the most significant research area of hyperspectral remote sensing image processing.Because hyperspectral remote sensing images have characteristics such as containing a large amount of data,being strongly connected among wave bands,having redundancy among wave bands,there are much difficulties and challenges in classifying hyperspectral remote sensing image by only using spectral data.Therefore,how to extract spatial spectrum information separately and combine the two for classification has become a key research direction in the field of hyperspectral image classification.Based on the semi-supervised learning method with small samples and the deep learning method with more samples,this paper has proposed two algorithms for how to extract effective spatial and spectral features.The specific content is as follows:Firstly,this paper has researched the commonly-used classification algorithms in the field of current hyperspectral remote sensing images,and has introduced the principles of commonly-used algorithms in the unsupervised classification methods,supervised classification methods and semi-supervised classification methods.At the same time,the paper also gives a detailed introduction to the basic theory and structure of convolutional neural network CNN in deep learning,and explains the basic structure flows of one-dimensional convolution,two-dimensional convolution and three-dimensional convolution.These studies have provided theoretical support and thoughts on the direction of improvement for the two hyperspectral image algorithms combined with spatial spectrum features proposed by this paper.Aiming at the situation that the small number of hyperspectral image samples leads to poor classification results,as well as the poor generalization ability of current hyperspectral image classification algorithms,this paper proposes a semi-supervised classification method of the hyperspectral remote sensing image based on small sample learning and combined with deep three-dimensional residual network.On the one hand,this method uses a deep threedimensional residual network to extract the spatial spectrum features in a hyperspectral image to reduce the uncertainties during labeling.On the other hand,it uses training samples to train the deep three-dimensional residual network to obtain a metric space,and then generalizes the learned metric space into the test data sets,and then combines the SVM algorithm to classify the hyperspectral images in the test sets.The experimental results show that this method can effectively alleviate the impact of the insufficient number of hyperspectral image samples on the classification effect,and provide excellent classification results in the case of small samples.In addition,this paper also shows that when the samples are sufficient,combining SVM to classify images can also achieve good classification accuracy,which further proves that the method has strong adaptability.In the field of deep learning,there has been problems that two-dimensional convolution cannot effectively use spectral features,and that three-dimensional convolution has difficulties in extracting spatial and spectral features at the same time.Based on these facts,this paper proposes a classification algorithm combined attention mechanism with mixture model CNNs.On the one hand,this algorithm combines three-dimensional convolution with two-dimensional convolution,so that the network can not only extract the spectral and spatial features of hyperspectral images at the same time,but also effectively overcomes the drawback that the traditional three-dimensional convolution is unable to provide a better classification effect for the reason that sample types have similar textures on many spectral bands.On the other hand,by introducing the attention mechanism and cascading the space and channel attention modules in the convolution process,not only the importance of channels in different features can be learned,but the importance of different positions in the same channel can also be learned,and higher weights are assigned to the feature channels and locations.Experimental results show that this method can obtain much more effective spatial and spectral joint features,and because the algorithm introduces an attention mechanism,more weights are assigned to valid features,which further improves the classification accuracy of the algorithm.