Research On Spectral Feature Selection Of Hyperspectral Remote Sensing Image Based On Mutual Information

With the gradual maturity of hyperspectral remote sensing technology,hyperspectral images have been widely used in many research fields,such as air pollution detection,crop chemical composition analysis,natural disaster prevention,etc.Hyperspectral remote sensing image data contains rich spectral and spatial information,which can accurately distinguish different types of land cover.However,in practical applications,due to the high frequency band sampling density,hyperspectral remote sensing image data has high dimensional frequency band number and large frequency band redundancy,which increases the complexity of data processing and causes the emergence of "Hughes" phenomenon and other problems.On the basis of making full use of the massive information provided by hyperspectral image data,improving the classification performance by reducing the dimension of the band has become one of the important research contents of hyperspectral remote sensing image processing.This paper has carried out research work on the optimal band selection of hyperspectral remote sensing images,and proposed a spectral feature selection algorithm of hyperspectral remote sensing images based on mutual information.The main research contents are as follows:(1)MCFS-UC spectral feature selection algorithm is proposed.This algorithm improves the CFS-UC feature selection algorithm,uses 3D mutual information to measure the redundancy of features and the correlation between features and classes,improves the accuracy of feature measurement,and can select better features.In addition,the algorithm introduces variance into the feature selection evaluation criteria to solve the problem of overestimated feature importance.Through this algorithm,the optimal feature combination with high correlation and low redundancy among features can be selected.This article conducted comparative experiments on the Indian Pines and Salinas datasets,and the experiments showed that the MCFS-UC algorithm outperformed the comparison model,and the classification accuracy was higher than the CFS-UC algorithm.(2)The MMRI-Boruta spectral feature selection algorithm is proposed.Based on the MRI feature algorithm,the algorithm introduces variance as a weight,taking into account individual differences,and reducing the impact of extreme values.In addition,MMRI-Boruta algorithm is a hybrid feature selection algorithm for hyperspectral remote sensing images,which combines the advantages of fast filtering algorithm,low time complexity and encapsulated algorithm to automatically determine the optimal feature subset size.This article also conducted comparative experiments on the Indian Pines and Salinas datasets,demonstrating that the classification performance of the MMRI Boruta algorithm is superior to the other five comparative models in most cases.(3)Design and implementation of feature selection system for hyperspectral remote sensing image based on Python.The system includes five modules: algorithm selection,data selection,parameter setting,classifier selection,drawing selection,and dimension reduction result display.The system has the advantages of simple and intuitive interface,simple operation and easy use.The development of this system is convenient for users to use the feature selection algorithm and other feature selection algorithms proposed in this paper to reduce the dimension of high-dimension remote sensing data.