Research On Hyperspectral Image Classification Algorithm Combined With Spatial Spectral Features

With the rapid development of imaging spectrometers,hyperspectral images can provide more ground object category information,and hyperspectral image processing technology has continued to make breakthroughs in recent years,which provides a basis for the further development of hyperspectral image classification technology.Hyperspectral image classification technology is an important research direction in the field of hyperspectral image processing.Because hyperspectral remote sensing images have the characteristics of large data volume,strong correlation between bands,and redundancy between bands,hyperspectral images are classified using only spectral data.Brings many difficulties and challenges.How to extract effective spatial information combined with spectral information for classification has become a research hotspot in hyperspectral image classification.In this paper,under the traditional method with small samples and the deep learning method with more samples,two algorithms are proposed for extracting effective spatial and spectral features.The specific contents are as follows:1.Aiming at the problem that the traditional local binary mode LBP is not accurate enough to obtain the spatial texture features,and the simple linear iterative SLIC algorithm obtains the super pixel segmentation map as a post-processing problem,a classification method combining improved LBP and super pixel decision is proposed.On the one hand,this method uses the newly proposed multi-directional cross LBP descriptor to obtain more effective spatial texture features in the form of statistical histograms,and combines the statistical histograms obtained under different window sizes,so that spatial texture features have different size ranges and higher accuracy.On the other hand,the improved SLIC algorithm is used to obtain a superpixel segmentation map that is more suitable for hyperspectral images,and the super-pixel segmentation map is used to guide and correct the classification results,which can effectively remove the scatter phenomenon and a certain degree of partial misclassification.Experimental results show that this method can obtain more effective spatial texture features and combine spectral features to improve classification accuracy.At the same time,the improved SLIC algorithm has more accurate edges,which can be used as a post-processing to further improve classification accuracy.2.Aiming at the problem that the two-dimensional convolution in the convolutional neural network CNN cannot effectively use the spectral features and the three-dimensional convolution while simultaneously extracting spatial and spectral features,a classification method of joint hybrid attention and dual-stream CNN is proposed.In the dual-stream architecture of the algorithm,the input data of one stream branch selects pixel blocks with larger spatial size,focusing more on extracting spectral-related spatial features,and the input data of the other stream branch selects pixels with larger spectral dimensions and smaller spatial sizes.Block,focuses more on extracting small local spatially related spectral features,and combining the features extracted from the two.At the same time,the attention mechanism is introduced in the dual-stream architecture,which not only combines the spatial attention mechanism with the input pixel block,pays more attention to the information related to the center pixel in the neighborhood,but also introduces the channel and space in the process of convolution.The hybrid attention mechanism can not only learn the importance of different feature channels,but also learn the importance of different locations of the same channel,and assign higher weights to effective feature channels and locations.Experimental results show that this method obtains more effective spatial and spectral joint features through a dual-stream structure,and the hybrid attention mechanism assigns more weight to effective features,which further improves the classification accuracy.