Band Selection For Hyperspectral Forest Monitoring Based On Image Quality Evaluation

Compared with multi-spectral,hyperspectral data have higher spectral resolution and richer spectral information.It is suitable to identify and classify tree species in forests,and have been widely used to monitor forest resources.In practical applications and current research,we need to select optimal bands from multiple,however,because of the large amount of data,it took long time and the efficiency was relatively low.In addition,after the band selection,the general researches lack consideration about quality of images and the analysis of the application effect.In this paper,we took the state-owned Gaofeng Forest Farm in Nanning City,Guangxi Zhuang Autonomous Region,as the research area,and used hyperspectral remote sensing data of AISA Eagle II.Combined with evaluation indicators of remote sensing image quality,we established a comprehensive evaluation model on image quality for the selection of hyperspectral data in forest monitoring.After that,we used the comprehensive evaluation model,the Adaptive Band Selection(ABS)and the Band Index(BI)to select the optimal bands of hyperspectral data,and combined the K-Means clustering method and the Optimun Index Factor(OIF)with the above three methods to study the optimal band combination and explore the advantages and disadvantages of each method.The main conclusions of this paper are as follows(1)Information Entropy,Signal-to-Noise Ratio,Coefficient of Variation,Contrast and Blur Degree based on Gray Level Co-occurrence Matrix(GLCM,it belongs to texture)have significant impact on remote sensing image quality.When we evaluated the quality of remote sensing images from different aspects,six evaluation indicators based on spectral information and seven evaluation indicators based on GLCM were selected.By comparing the correlation between any two indicators,Information Entropy,Signal-to-Noise Ratio,Coefficient of Variation,Contrast and Blur Degree based on GLCM stood out to participate in the construction of comprehensive evaluation model for their low correlation.It indicates that they can comprehensively describe sharpness,the degree of noise influence,and the amount of information of images(2)We obtained the comprehensive evaluation model(CEM)for evaluating the quality of hyperspectral images.The coefficient of determination of model reached 0.707,and the Root Mean Square Error(RMSE),Estimation Accuracy(EA)and Relative Error(RE)were 0.029,99.94%,0.42%.In addition,it had effect on other images,indicating that the model has a strong ability to evaluate the image and the effect is significant.The top 10 bands selected by CEM were compared with the top 10 bands selected by Band Index(BI)and the top 10 bands selected by Adaptive Band Selection(ABS),and we calculated the overall classification of these bands.The results shows that the precision of CEM is highest and the overall effect is much better in forest monitoring.(3)We used K-Means clustering to solve the problem of inter-band correlation which was not considered by CEM.When the number of bands in the band combination is three,the optimal band combination under the method of K-Means&CEM,BI&OIF,ABS&OIF were 17-55-84,41-62-11 and 58-99-110.By comparing the average correlation in band combination,index value of OIF and classification accuracy of the optimal band combinations above,the band combination selected by K-Means&CEM ranked the first.(4)Generally,when the number of bands in the band combination is 3 to 10,with the increase of the number of bands in the band combination,the classification accuracy and the correlation among the bands in the combination of the band combination selected by K-Means&CEM,BI&OIF and ABS&OIF are increasing.In addition,K-Means&CEM has the highest classification accuracy,and the correlation among the bands in the combination is the lowest.