| The content of nitrogen and phosphorus in soil that has a great influence on crops.Fast,real-time,large-area monitoring of the content of nitrogen and phosphorus in the soil is of great significance for precision agriculture and soil chemical pollution.With the rapid development of hyperspectral technology,It has become one of the important methods to estimate soil trace elements through hyperspectral inversion.In this paper,soil data were used.Using ASD spectrum instrument to collect spectral reflectance of soil under the condition of the laboratory experiments,in combination with the determination of the chemical data of FD、CR、SD and so on eight different kinds of spectral filtering characteristics of soil total nitrogen and total phosphorus,with characteristic bands for the independent variable to establish and PLSR,SVR to BPNN neural network inversion model.Through the analysis of the inversion model,the following conclusions are drawn:(1)The TN content was classified into three orders of magnitude according to the content statistics of samples and according to 800～1000mg,1000～1200mg and1200～1400mg,corresponding to N0 N1 and N2 respectively.The reflectivity corresponding to the whole spectral band was not linearly related to the TN content.The reflectance of 400～800nm band is close to each other.The reflectivity corresponding to N2 in the band interval between 800nm and 1900nm was higher than that of N0 and N1.The TN content is proportional to the reflectivity N2>N1>N0in the band of 1900nm～2400nm.According to the different sample contents,TP can be divided into three orders of magnitude ranging from 500 to 700mg,700 to 800mg and 800 to 900mg,named after P0、P1 and P2respectively.In the band range of400nm to 600nm,the reflectivity of P1 is higher than that of P0 and P2,and the reflectivity of P0 and P2 coincide.In the band range of 600～2400nm,P1>P2>P0,and the reflectivity difference is large(2)Different soil spectral reflectance data and its corresponding to the TP content of the correlation analysis(a total of eight different spectral transform)TN and TP in the characteristic band,including TN characteristic bands on:800～1100 nm,1300～1800 nm and 2100～2300 nm,the characteristic band mainly concentrated in the more than 1000 nm,focused on the characteristics of TP band:407～503 nm,1000～1300 nm,1400～1800 nm and 2100～2300 nm.(3)The inversion model of soil characteristic spectrum and TN and TP was constructed based on eight different spectral transform screening feature ba nds as independent variables and combined with PLSR,BPNN and SVR algori thms.The best inversion model of TN:the SVR inversion model with RBF as the kernel function and the feature bands screened by SD transformation as the independent variables,verified that R2=0.83,RMSE=0.97mg/kg;The PLSR inver sion model with the features screened by FD transform as independent variable s verified that R~2=0.71,RMSE=1.21mg/kg.Using the transfer layer function as T ansig,the training function as Traincgf,and the feature band of CR transform screening as the independent variable BPNN inversion model,the verification R~2=0.73,RMSE=1.13mg/kg.The best model of TP:the SVR inversion model wit h RBF as the kernel function and the feature bands screened by FD spectral tr ansformation as independent variables,verified R~2=0.77 and RMSE=0.13mg/kg.PLSR inversion model with the feature bands screened by SD transformation a s independent variables verified R~2=0.68,RMSE=0.21mg/kg.Taking the transfer f unction as Purelin,the training function as Traincgb,and the feature band scre ened by SD transformation as the independent variable,the BPNN inversion m odel verified that R~2=0.84,RMSE=0.09mg/kg.(4)TN model inversion,SVR>BPNN>PLSR;In terms of the inversion effect of TP model,BPNN>SVR>PLSR;Generally speaking,the inversion effect of nonlinear model is better than that of linear model.In terms of TN and TP inversion results,the verification accuracy of the linear model PLSR is about 0.7,and RMSE is larger than that of the nonlinear model RMSE. |
PDF Full Text Request