Hyperspectral Characteristics And Quantitative Estimation Of Soil Properties In Red Soil Region Of Southern Jiangxi

The problem of soil degradation in the red soil region has a long history.It is an important prerequisite for soil quality monitoring and soil ecological restoration to quickly and effectively obtain soil physical and chemical properties.Nowadays,soil hyperspectral technology has become an important choice for rapid acquisition of soil physicochemical properties rely on its fast speed,low cost,and no damage.However,in the case of diverse soil types and different properties,it is still difficult to achieve effective prediction of soil physical and chemical properties.In this study,673 valid field soil samples were collected from Xinfeng,Anyuan and Xunwu counties in Ganzhou city of Jiangxi province.According to the physicochemical properties and hyperspectral data of soil acquired indoors,the spectral characteristics of different soil types were analyzed.Based on the synthesis of soil composition and attributes and the continuity of spectrum,an attempt was made for remove interference factors to explore a spectral data preprocessing method.Subsequently,each preprocessing was performed on the original spectral data,and the differences in soil spectral properties under different preprocessing were described.The correlation characteristics of soil clay content(Clay),organic matter content(SOM),total nitrogen content(TN),total carbon content(TC),p H value(p H)and total copper content(Cu)with different preprocessing methods in full spectrum,feature spectrum and spectral index were further analyzed.Hereafter,the feature spectrum and sensitive spectral index combinations were selectd by correlation coefficient and variable importance in projection(VIP).Finally,the performance of the prediction models constructed by different methods were compared,and the optimal spectral prediction model were selected.The main conclusions were as follows:(1)The spectral reflectance curves of red soil,yellow soil,paddy soil and purple soil are quite different in the visible-near infrared band.The spectral reflectance morphology of red soil and yellow soil is similar in visible part,and differences in near-infrared part.Specifically,the spectral reflectance curves of red soil and yellow soil are similar at 800-1400 nm and 1400-1900 nm,but the reflectivity of red soil is generally higher than that of yellow soil.while at 2100-2400 nm,the morphology of them is opposite.The reflectance spectrum of paddy soil showed a gentle upward trend in general,and the spectral reflectance at 1800-2400 nm was similar with that of yellow soil and purple soil.Purple soil posed a special reflection spectrum curve.The morphology at 400-1100 nm was similar with that of red soil,and the reflectance at 1100-1400 nm decreased slowly.The trend of reflectance spectra of purple soil in the range of 1400-2400 nm is similar with that of yellow soil,but showed a lower reflectance.(2)The continuous wavelet transform(CWT)showed no obvious effect on the correlation of soil spectrum in full spectrum.Both 1.5th fractional order derivative(FOD(1.5))and interference factor removal(IFR)can improve the correlation between soil SOM,TN and TC contents with soil reflectance spectrum.However,the correlation of the original spectrum(R)is significantly improved by IFR,with the improvement magnitudes of 0.23,0.27 and 0.23,respectively.In addition,none of CWT,FOD(1.5)and IFR could improve the correlation of soil reflectance spectrum with soil p H and Cu content.In terms of spectral indices,CWT could not improve the correlation between spectral index and soil properties compared with the difference index(DI),ratio index(RI)and normalized difference index(NDI)in R.FOD(1.5)and IFR preprocessing could enhance the correlation between spectral index and Clay,SOM,TN and TC.In detail,FOD(1.5)can improve the correlation between DI and them,while IFR enhances the correlation between RI and NDI more significantly.The results also showed that CWT,FOD(1.5)and IFR reduced the correlations of the three spectral indices with soil p H and Cu content.(3)The feature spectrum was selected from the full spectrum(401-2400 nm)by VIP value,and the results showed that the number of feature bands selected by the six soil properties under the four pretreatments was:R=CWT>IFR>FOD(0.5).The number and positions of feature bands extracted by CWT and R are similar,which showed that the effect of CWT preprocessing is not obvious in this study.Based on the VIP value,the range of the overlapping feature bands of the six soil properties under the four pretreatments was further clarified,and the relatively fixed feature band positions were explored hereafter.The results show that the overlapping regions of the feature bands of Clay content are located near 1350-1450 nm,1850-1930 nm and 2120-2300 nm.The overlapping regions of the feature bands of SOM content,TN content and TC content are concentrated around 1400 nm,1900 nm and 2250 nm.The overlapping regions of the feature bands for p H and Cu content are similar,located at 430 nm,1400 nm,1900 nm and 2200 nm both.According to the overlapping areas of the feature spectra of the six soil properties,they all close to the absorption position of soil moisture,showing that the selection of feature bands is still affected by soil moisture in varying degrees.(4)The prediction ability of PLSR,ELM and PLSR-ELM models for six soil properties,including soil Clay content,SOM content,TN content,TC content,p H value and Cu content,were compared in three aspects:full spectrum,feature spectrum and spectral index.It is shown that the PLSR-ELM model kept a highest prediction accuracy for Clay content,SOM content,TN content,TC content and p H value,while the PLSR model showed the highest accuracy for Cu content prediction.Concretely,the spectral model of SOM and TC content showed the highest prediction accuracy,with R~2 of validation dataset is 0.75 and 0.70through the approach of sensitive spectral index,IFR and PLSR-ELM.The spectral model of Clay and TN content showed the highest prediction accuracy,with R~2 of validation dataestis 0.66 and 0.62 by full spectrum,IFR and PLSR-ELM.With the feature band,IFR and PLSR-ELM,the spectral model of p H value showed the highest prediction accuracy with the R~2 of validation set is 0.48.With full spectrum,FOD(1.5)and PLSR,the prediction accuracy of the spectral model of Cu content is the best,and the R~2 of the validation set is0.51.However,the validation accuracy of soil p H and Cu content models are limited,the effective spectral prediction model for p H and Cu content could not be obtained in this study.Overall,except for the spectral prediction model of Cu content,the best modeling approaches for the other five soil properties both included IFR preprocessing and PLSR-ELM method.It indicated that the preprocessing method of interference factor removal(IFR)that established based on the synthesis and removal of interference factors can improve the spectral prediction accuracy of soil Clay content,SOM content,TN content,TC content,and p H value,which is beneficial for their spectral predictions.At the same time,the composite model PLSR-ELM formed by combining PLSR and ELM posed a better ability in spectral prediction in this study.