| Soil organic matter(SOM)content and soil total nitrogen(STN)content are important indicators of soil quality and fertility,as well as important sources of nutrients to support crop growth.Remote sensing technology is a fast way to acquire soil property information in the field,and the hyperspectral imager carried by unmanned aerial vehicle(UAV)has the advantages of high spatial resolution,low cost and flexibility and freedom,which provides an alternative data source for rapid quantitative monitoring of soil nutrients.However,there are problems of high spectral noise and redundant spectral information in UAV hyperspectral images.Therefore,this paper takes agricultural soils located in the black soil region of northeast China as the research object,and achieves higher accuracy of soil nutrients estimation and mapping based on UAV hyperspectral images in cultivated land by studying spectral denoising,spectral feature selection and inversion modeling methods.The research contents and main conclusions of this paper are as follows.(1)To address the problem of high spectral noise in hyperspectral images caused by lighting conditions and other factors,this paper compares Multiple Scattering Correction(MSC),The Standard Normal Variate transformation(SNV),the first derivate(FD),and the second derivate(SD)to investigate the suitable methods for spectral noise removal from hyperspectral remote sensing images.The results show that MSC improves the signal-to-noise ratio from 31 dB to 160 dB,and increases the absolute values of the maximum correlation coefficients between the spectral reflectance and SOM and STN from 0.27 and 0.28 to 0.51 and 0.59,respectively.Compared with other methods,MSC effectively reduces the noise of the original spectra and enhances the relationship between the spectral reflectance and soil properties.(2)To address the problem of high dimensionality and data redundancy of hyperspectral data,a combination of Competitive Adaptive Reweighted Sampling(CARS)and Successive Projections Algorithm(SPA)is proposed and compared with the traditional correlation coefficient method.The results show that the redundancy of the feature band set screened by using only the CARS method or correlation coefficient method is large,while using only the SPA method can effectively reduce the inter-band redundancy but lose the useful information related to the target attributes.The combined CARS-SPA algorithm can further reduce the inter-band redundancy while retaining as much useful information as possible.The inter-band redundancy is reduced from 112 bands to 24(SOM)and 22(STN)effective feature bands to introduce suitable predictors for inversion modeling.(3)Using particle swarm optimization(PSO)to optimize the Extreme Learning Machine(ELM),we use the feature bands of hyperspectral image data as input variables to construct the prediction models of SOM and STN,and compare with the traditional linear regression,support vector machine(SVM)and Back Propagation Neural Network(BPNN).The results show that for SOM prediction,the PSO-ELM machine learning model based on the CARS-SPA feature band set has the highest accuracy,with the validation set R2=0.73,RMSE=2.63 g·kg-1,and RPD=1.91;for STN prediction,the three machine learning models including PSO-ELM,SVM and BPNN based on the CARS-SPA feature band set models all achieve good prediction performance,with the validation set R2 of about 0.63,RMSE of about 0.22 g·kg-1,and RPD of about 1.53.Finally,the optimal inversion models of SOM and STN were selected to realize the spatial inversion of soil nutrients and analyze their spatial distribution characteristics.This paper provides a reference for the application of unmanned airborne hyperspectral technology in quantitative soil nutrient inversion,which can provide scientific technical guidance for implementation of precision agriculture and is important for scientific fertilizer allocation for field management as well as soil environmental quality monitoring and assessment. |
PDF Full Text Request