Research On Classification Method Of Hyperspectral Images Based On Convolution Neural Network And Integrated Learning

With the continuous development of science,remote sensing technology is gradually infiltrating into the national economy,ecological environment,national defense science and technology,and plays an irreplaceable role.Hyperspectral image classification technology is one of the main research directions of remote sensing science,and its purpose is to determine the feature types in remote sensing images.Higher classification accuracy helps people better understand the distribution of features,so it is important to seek image classification algorithms with high classification accuracy.At present,there are many classification algorithms.However,due to the particularity of hyperspectral data,there are still many problems in hyperspectral image classification algorithms.Based on the problems of current classification algorithms,two hyperspectral image classification algorithms are proposed.The research work of this paper is described as follows:Aiming at the problem of using only spectral information in traditional hyperspectral image classification,this paper proposes a method for extracting spatial-spectral information,which effectively collects joint information of spectral space.Considering the outstanding performance of the current tree integration method in the data science competition,this paper uses the Light Gradient Boosting Machine(LightGBM)algorithm for hyperspectral image classification,and then combines the space-spectrum information extraction method to propose a new Hyperspectral image classification algorithm(KC-LGM).Experiments show that the algorithm can greatly improve the classification accuracy.The traditional hyperspectral image space-spectral classification algorithm relies heavily on domain knowledge,can only obtain "shallow" layer features,and requires artificial synthesis features;hyperspectral image classification algorithm based on 1D convolutional neural network can not utilize spatial information of data;based on 2D The hyperspectral image classification algorithm of convolutional neural networks requires the convolution kernel thickness to be equal to the number of spectral bands.Therefore,the original data must be dimensionally reduced,which will lose a certain degree of spectral information.In view of this,considering the 3D convolutional neural network for 3D convolution,it can improve the defects of 2D convolutional convolution kernel over-thickness and 1D convolution can not extract spatial information.This paper proposes a new hyperspectral image classification algorithm based on 3D convolutional neural network.The algorithm constructs a general hyperspectral image classification framework,which uses a 3D filter to simultaneously perform convolution operations from three directions,and by setting an appropriate convolution step,a part of the convolutional layer can be pooled at the same time as the convolution operation which has improved the efficiency of the network.Experiments show the effectiveness of the proposed algorithm.