The Research Of Prediction Model Of Soil Organic Matter Content Based On Hyperspectral Data In Different Sampling Depth

Soil organic matter is an important factor in soil fertility and crop production.It can improve the physical properties of soil,facilitate microbial activities such as flora,and promote the decomposition of nutrients in the soil to ensure the growth and development of plants.In order to adapt to the actual needs of precision agriculture,mining high-efficiency information based on hyperspectral technology has become an important direction for predicting soil organic matter content quickly,conveniently and accurately.In this paper,the experimental soil of Yancheng City,Jiangsu Province is taken as the research object.Three data sets are obtained in the field:sampling depth A(0-20 cm),sampling depth B(20-40 cm)and sampling depth C(40-60 cm).Based on the hyperspectral technique combined with the modeling method in chemometrics,multiple prediction models of soil organic matter content are studied.The main research contents and conclusions are as follows:(1)By analyzing the characteristics of the spectral reflectance curve,it is found that the spectral reflectance at 400-1000nm increases with the increase of the wavelength,the increase of the reflectance at 400-600nm is more obvious,and the increase of the reflectance at 600-1000nm slows down gradually.However,overall,the spectral reflectance of the three sampling depths are significantly different,and the sampling depth A<sampling depth B<sampling depth C.(2)In the pretreated stage,the comparison uses Savitzky-Golay smoothing(SG),First order differential(1stD),Second order differential(2ndD),Multiplicative Scatter Correction(MSC),Standard Normal Variation(SNV)effects of five single pre-processing and four combined pre-processing(SG+1stD,SG+2ndD,SG+MSC,SG+SNV).The results show that SG smoothing can preserve the original curve shape relatively well,and the effect is better than the original spectrum and other pretreated methods.At the same time,the predictors of the above nine pretreated methods are still low,indicating that there may be more serious multicollinearity problems between the wavelength variables,and it is necessary to screen the characteristic wavelengths.(3)During the characteristic wavelength selection period,Uninformative Variable Elimination(UVE),Bootstrapping Soft Shrinkage(BOSS),Interval combination optimization(ICO),and Interval Variable Iterative Space Shrinkage Approach-Successive Projections Algorithm(IVISSA-SPA)are used.On the data set of sampling depth A,sampling depth B and sampling depth C,the IVISSA algorithm can effectively filter out the redundant variables in serial connection with the SPA algorithm so that the number of characteristic wavelengths reduce from 232 to 17,103 to 23,182 to 31 respectively,which decrease by 92.67%,77.67%and 82.97%respectively;R2 increase from 0.7058 to 0.8002,0.5106 to 0.6229,0.5781 to 0.6103 respectively,which increase by 13.37%,21.99%and 5.57%respectively;RMSEP reduce from 2.9550 to 2.1255,2.2304 to 1.6890,1.2365 to 0.8480,decreased by 28.07%,24.27%,31.42%;RPD increase from 1.0849 to 1.5084,0.8815 to 1.1640,0.8209 to 1.1969 respectively,increased by 39.04%,32.05%,45.80%.The results show that the IVISSA-SPA algorithm can filter out the collinear variables effectively,which is beneficial to reduce the complexity of the model and improve the computational efficiency.Compared with other three characteristic wavelength selection algorithms,IVISSASPA>ICO>BOSS>UVE.(4)Partial Least Square Regression(PLSR),Support Vector Regression(SVR),BP neural network,Particle Swarm Optimization(PSO),Beetle Antennae Search(BAS)and Modified Beetle Antennae Search(MBAS)are used in regression modeling,which optimize the prediction effects of multiple models constructed by BP neural network.The results show that MBAS can optimize the initial weight threshold of BP neural network,improve network performance and avoid the problem of the network to fall into local optimum caused by the random initialization of BP neural network.Among all the prediction models,the IVISSASPA-MBAS-BP model has the best indicators.The R2 of the sampling depth A data set is 0.8896,the RMSEP is 1.2857,while the RPD is 2.4936.R2 on the sample depth B data set is 0.7284,1.0245 on RMSEP and 1.9190 on RPD.R2 on the sample depth C data set is 0.7250,RMSEP is 0.5492 and RPD is 1.8481.