Inversion Of Citrus Fruit Leaf Area Index Based On UAV Remote Sensing

Citrus,as the second largest fruit variety in China,plays an important role in China’s agricultural import and export trade,and its output is directly related to the volume of citrus import and export.However,the growth of citrus plants directly determines its yield.Now,Leaf Area Index(LAI),an evaluation index of crop growth,can be used to monitor citrus growth.In recent years,with the development of UAV remote sensing technology,it has been gradually applied to the rapid and non-destructive estimation of plant LAI.In this paper,Pengyu Brothers Citrus Experimental Base located in Gongcheng County,Guilin City,Guangxi Province was selected as the research area.In the morning of September7,2022,multi-spectral images and hyperspectral images of citrus canopy were collected by UAV,and measured LAI data were collected at the same time.Then the multispectral features and hyperspectral features used to establish the inversion model were extracted respectively,and the feature variables were screened by feature selection.Then,a single algorithm(partial least squares regression,support vector regression,random forest regression,BP neural network)and the Stacking integrated learning algorithm were used to establish LAI inversion models,and the inversion effects of different stacking combinations were compared and analyzed.The main research results of this paper are as follows:(1)When using multi-spectral data for LAI inversion of citrus fruit trees,inversion accuracy of texture feature and spectral index feature was higher than that of multi-spectral reflectance inversion,and the inversion effect of combining texture feature and spectral index algorithm was better than that of stacking integrated learning algorithm,R2 and RMSE of texture feature were 0.573 and 0.471,respectively.The R2 and RMSE of spectral index were0.687 and 0.451,respectively.(2)When using hyperspectral data for LAI inversion of citrus fruit trees,traditional spectral transform,continuous wavelet transform and spectral index can improve the accuracy of inversion model.The inversion accuracy of the model based on the traditional spectral transform features combined with a single algorithm is higher,and the highest accuracy is the multiple scattering correction spectrum combined with SVR,R2 and RMSE are 0.759 and0.286,respectively.The inversion accuracy of the model built with spectral index features and Stacking ensemble learning algorithm is high.The highest accuracy is the normalized spectral index stacking model,with R2 and RMSE being 0.932 and 0.194,respectively.The inversion accuracy of the model built by continuous wavelet transform spectral features and Stacking integrated learning algorithm is high.The highest accuracy is the sym2 combined stacking model,and R2 and RMSE are 0.903 and 0.270,respectively.(3)By comparing the inversion accuracy of different data sources combined with different models,it can be seen that the inversion accuracy of LAI inversion model based on hyperspectral data is higher than that of the model based on multi-spectral data.The inversion accuracy of multi-spectral data is concentrated between 0.388-0.687,and that of hyperspectral data is concentrated between 0.542 and 0.932.The model performance of a single algorithm from best to worst is BP neural network,support vector regression,random forest regression,partial least squares regression.In contrast to the inversion accuracy of single algorithm and Stacking integrated learning algorithm,the inversion accuracy of support vector regression and BP neural network in single algorithm is higher than the stacking model when modeling with multi-spectral data features.When modeling with hyperspectral data features,the inversion precision of traditional spectral transform data and continuous wavelet transform spectra combined with support vector regression and BP neural network was higher,and the inversion precision of spectral index combined with Stacking model was higher.