Study On Mining Soil Organic Carbon Signal And Improming Its Prediction Accuracy From Vis-NIR Spectral

Due to the role of soil organic carbon(SOC)in soil quality assessment,there is an increasing demand for data on soil organic carbon.Moreover,soil organic carbon content is key property related to soil physical,chemical and biological fertility.Indeed,it is an important determinant of soil sustainable development.Visible and near infrared reflectance spectroscopy(Vis-NIR)is widely used as a rapid and cost effective method to quantitatively infer SOC.While conceptually attractive,a primary challenge for using soil spectral in field application is the wide range of soil moisture that will be encountered,and field samples will be in quite different conditions in terms of moisture.Some effective method should be developed to improve predict accuracy of SOC content.In present study,a total of 70 soil samples about paddy soil were collected by aluminum boxes to remain the field conditions.We used the following steps to process reflectance:fractional derivative(FD),discrete wavelet packet transformation(DWPT)and location correlation maximization(LCM).Partial least squares regrelession(PLSR)moder was used to predict organic carbon content.Finally,9 soil moisture levels are derived to understand the fitness for use of those series of methods.The main conclusions are as following:(1)Some interim signal can be discovered by extending the integer calculus to fraction derivative,due to the advantages of memory and nonlocality.Derivative could evidently increase the significant correlated with SOC content(0.6-order>1st-order>2nd-order),the number of highly significantly correlated bands reaching the maximum at 0.6-order,the accuracy of SOC content predict model(R2val=0.693,RMSEV=1.952 g/kg,RPD=1.85)have a significantly improve compared with original reflectance(R2val=0.663,RMSEV=2.045 g/kg,RPD=1.77).(2)The wavelet packet transform of the 0.6-order derivative reflectance spectra can effectively de-noising,witch one was amplified by derivative or transformed from original spectra,and reaching the maximum correlation coefficient at level 6;combination of derivative and wavelet transform,de-nosing of the 0.6-order derivative reflectance had a best model accuracy(R2val=0.727,RMSEV=1.840 g/kg,RPD=1.97)of predict SOC contend than derivative spectra(R2val =0.693,RMSE=1.952 g/kg,RPD=1.85)or wavelet packet transform spectra(R2val=0.663,RMSEV=2.044 g/kg,RPD=1.77).(3)Local correlation maximization method has great potential to monitor SOC signal when reduces noise while retaining as much soil organic carbon information as possible,derived a great model(R2val =0.781,RMSEV=1.679 g/kg,RPD=2.17)in predict SOC content,and provides a satisfying robustness(R2val =0.797,RMSEV=1.660 g/kg,RPD=2.18)at the same time from bagging test.There are 3 pink values located in visible,1450 and 1950 nm respectively,according to variable in the projection(VIP)analysis,witch illustrate that LCM can't remove the effects of soil water in reflectance exhaustively.(4)Soil organic carbon content estimation model was affected by soil moisture content largely,especially when soil moisture content from 50 to 250 g/kg.when soil moisture content increased from 250 to 450 g/kg,SOC spectral characteristics was covered by soil moisture spectral characteristics,it was not suitable to be used for SOC content estimation.LCM Could considered as an effective preprocessing method to improve SOC content modeling accuracy in different soil moisture level,especially in the soil moisture content from 250 to 450 g/kg,and have potential to remove the influence of soil moisture when soil moisture content from 250 to 450 g/kg.