Research On Hyperspectral Image Classification Method Based On Combined Spatial-spectral Features And Machine Learnin

Hyperspectral images are widely used in the fields of resource exploration,urban planning and military because they contain rich spectral and spatial information and have the characteristics of unity of map and high spectral resolution.As an important means for human to understand the world,hyperspectral image classification technology is one of the key technologies in the field of hyperspectral image processing.Aiming at the problems of high dimensionality of hyperspectral image data,redundant information,and poor classification results due to inadequate extraction and utilization of spatial spectral features in hyperspectral image classification tasks,this paper combines band selection methods,joint spatial spectral features,and uses integrated learning and deep learning techniques to study hyperspectral image classification methods.The main work of this paper includes:(1)Aiming at the problem that insufficient feature extraction of hyperspectral images leading to poor classification results,a hyperspectral image classification method using Light GBM and band selection is proposed.Firstly,the Wasserstein distance improved m RMR algorithm is proposed to select the optimal subset of bands with low redundancy among bands and high correlation between bands and feature classes from the original bands through band selection,and the quadratic simplification of band dimensions and extraction of spectral features are achieved by principal component analysis for the optimal subset of bands;secondly,the grayscale,Gabor features are extracted at multiple scales using Grey Level Co-generation Matrix and 2D Gabor filter for the principal component features;finally,the hyperspectral image classification is performed by using Light GBM classifier after combining the spatial-spectral features.The method is used for classification experiments on WHU-Hi-Long Kou,WHU-Hi-Han Chuan and WHU-Hi-Hong Hu datasets,and the experimental results prove that this method has better classification result compared with traditional methods.(2)Aiming at the problem that the traditional feature extraction methods for hyperspectral images neglect deep-level feature mining leading to low classification accuracy,a hyperspectral image classification method using Light GBM and convolutional neural network is proposed.Firstly,the raw data are preprocessed by band selection and principal component analysis;secondly,a three-dimensional convolutional neural network with the introduction of a convolutional attention mechanism module is constructed to extract deep-level spatial-spectral features,while a Grey Level Co-generation Matrix is used to extract grayscale features at multiple scales;finally,the hyperspectral image classification is performed by combining the deep and shallow spatial-spectral features using the Light GBM classifier.The experimental results on three hyperspectral image datasets demonstrate that this method outperforms other methods.(3)Aiming at the problem that inadequate utilization of spatial-spectral information and high performance requirements of computing devices in the hyperspectral image classification method by 3D convolutional neural networks,a hybrid convolutional neural network hyperspectral image classification method incorporating multiple attention mechanisms is proposed.Firstly,perform band selection and principal component analysis on the original data to obtain the principal component features,and resample the principal component features to expand the sample size;secondly,construct a hybrid convolutional neural network model containing one-dimensional,two-dimensional and three-dimensional convolutional neural network modules and a two-branch triple attention mechanism module,extract the spatial-spectral features by the three-dimensional convolutional neural network module,introduce the two-branch triple attention mechanism to achieve cross-dimensional information interaction and capture important features,the spatial and spectral features are enhanced by using two-dimensional and one-dimensional convolutional neural network modules,respectively,while reducing the complexity of the network model;finally,the spatial-spectral features are input into the softmax layer for classification.The experimental results on three hyperspectral image datasets demonstrate that this network has better classification ability compared with other networks.