Pixel-Level Hyperspectral Image Classification Based On Joint Self-Organizing Map Network Dimensionality Reduction And Supervised Learning

In recent years,the rapid development of remote sensing detection technology has significantly improved the spectral resolution and spatial resolution of hyperspectral image data,and has provided better support for the detection of ground targets.The classification and recognition of hyperspectral image data is currently a research hotspot in the field of remote sensing.As the dimensionality of spectral information in hyperspectral remote sensing images is getting higher and higher,and the information in the spatial structure is becoming more and more complex,it also brings great difficulties to the classification of ground objects.How to effectively reduce the dimensionality and specifications of hyperspectral image data and meet the fine-grained classification and prediction of ground objects has become an urgent research topic.Aiming at this problem,a pixel-level classification method and implementation process of hyperspectral image data based on data protocol are proposed.Based on the use of self-organization feature map(SOM)to cluster the hyperspectral image data,the spatial feature extraction of the ground objects is performed,and the machine learning and deep learning classifiers are combined for pixel-level recognition.First,the Principal Component Analysis(PCA)method is used to compress the dimensions in the spectral domain of hyperspectral image data,which realizes data compression while ensuring the information recognition effect.Then,SOM models of different granularities are used to perform unsupervised clustering of PCA dimensionality reduction data to achieve pixel-level data reduction and dimensional compression of hyperspectral images.Secondly,in the multi-granularity SOM topology map,the neighborhood frequency features of the pixels are extracted to realize the feature expression of the target pixels.On the one hand,a variety of machine learning methods are used for feature recognition,and the feature importance of the corresponding component of each feature is obtained.On the other hand,whether the attention-based convolutional neural network is used to obtain the attention information of the corresponding component of each ground object.Finally,according to the statistical learning principle,the information in its spatial neighborhood is counted in the multi-granular SOM topology structure diagram.And the frequency features extracted in the multi-granular SOM topology map select different machine learning models and the importance of features established by deep learning models for quantitative analysis.In the three methods of statistical model,machine learning model,and spatial attention,the component information in the hyperspectral image data is accurately disassembled to obtain the pixel-level classification results of the remote sensing image.Through the component information existing in the SOM in different granularities,the information existing between different features is fine-grained,and the topological structure information existing in different real features and the importance of the machine learning model and the attention in the deep learning model are analyzed.The mechanism enhances the interpretation of ingredient information.Using the hyperspectral image data of Matiwan Village in Xiong’an,the information about the features of the features is obtained and the pixel-level feature classification effect is improved.The interpretation of ground feature information through PCA dimensionality reduction and SOM data protocol provides new research ideas for ground feature classification and quantitative prediction.