Research On Hyperspectral Object Classification System Based On Feature Fusio

Hyperspectral images(HSI)are an important part of the remote sensing field,where each pixel can be seen as a high-dimensional vector with the number of dimensions corresponding to the spectral reflectance at a particular wavelength.Because HSI carries both spatial and spectral information,whose data is highly dimensional and redundant.Many traditional machine learning methods are difficult to deal with such complex data,and although researchers have introduced deeplearning-based algorithms to improve the classification accuracy of HSI,they also pose the problem of multiplying the computational effort.To extend fully automated HSI classification algorithms to practical applications,it is necessary to develop fully automated classification algorithms,which can balance computational efficiency and classification accuracy.This dissertation proposes a hybrid dual-channel structure using two fusion mechanisms and two attention modules to achieve accurate and efficient HSI classification.The experimental results on three widely used public HSI datasets demonstrate the performance of our model.In addition,a web-based software system is designed the use of existing software framework technologies to move the HSI classification algorithm towards practical applications.The main work of the thesis is as follows.(1)Based on a thorough study of 3-D and 2-D convolutional networks and fusion mechanisms,a hybrid dual-channel structure based on feature fusion mechanisms is proposed.In the network,the 2-D convolutional channel extracts pure spatial features and the 3-D convolutional channel extracts joint spectral-spatial features;the two channels extract features separately to avoid interference between the different features.In order to make full use of these different features,two fusion strategies are designed to enhance the dual-channel structure.The real-time fusion is applied to each convolutional layer between the two channels,enhancing the information interaction within the network.The stacked fusion strategy uses adaptive weights to enhance useful features and suppress useless features in order to adaptively adjust the impact of the two channels on the final classification result.(2)In addition,to enhance the feature representation capability of the network,the Spectral-Spatial Reconstruction Attention Module(SSRAM)and the Adaptive Channel Attention Module(ACAM)are designed specifically for the dual-channel structure after careful study of the current mainstream attention mechanisms.The SSRAM based on feature reconstruction is used in the 3-D convolutional channel to improve the sensitivity of spectral spatial information,while the ACAM based on multi-scale adaptive weighted statistics can selectively emphasize information features to enable smoother fusion operations in the 2-D convolutional channel.(3)Finally,based on the above fully automated classification algorithm,this dissertation designs and implements a web-based framework for a hyperspectral remote sensing image classification software system using technologies such as Django,Vue and Ajax.The system has the functions of account management,image classification,classification result analysis and historical data management,which can meet the basic needs of scientific research and production practice activities.The test results on three computers with different configurations show that the software system can achieve the expected functions and run stably.