Rapid Determination Of Soil Class Based On Proximal Soil Sensing

Soil is a vital component of the Earth’s critical zone.Soil classification is the basis for understanding soil.Traditional soil classification methods are not only complex and time-consuming,but also challenging to meet the current needs of people for high-precision and high-efficiency soil information.Spectral technology has many advantages that conventional laboratory analysis cannot match,such as being efficient,fast,convenient,non-destructive,and inexpensive,is widely applied in pedology.Soil spectral classification has been proven to assist in evaluating traditional classification systems and exist as an independent basis for classification.Previous studies have mainly focused on using a single spectrum to predict soil classes,and most use visible near-infrared(vis–NIR)spectroscopy.This study compared the prediction abilities of vis–NIR spectroscopy,mid-infrared(MIR)spectroscopy,and their fusion spectra for soil classification.In the global soil spectral database,60 typical soil profile data were selected for a total of four soil classes,and four spectral data,vis–NIR,MIR,and their simple combination(vis–NIR–MIR)and Outer Product Analysis(OPA),were investigated.Soil classification was conducted using K-means cluster,partial least squares discriminant analysis(PLSDA),and random forest(RF).The results show that K-means clustering is difficult to distinguish any soil type from the four spectral data.This is mainly due to the significant differences in data space,the high level of sample classification units,and the common phenomenon of “same spectral but the different components”.The classification verification results show that the performance of the nonlinear model RF(with an average classification accuracy of85.5% and an average Kappa coefficient of 0.81)is significantly superior to the linear model PLSDA(with an average classification accuracy of 68.1% and an average Kappa coefficient of 0.57)in the four spectral data.The classification model of the MIR spectra is superior to the classification model of the vis–NIR spectra.The minimum accuracy of the classification model based on the MIR spectra(MIR–PLSDA)is 71.4%,and the minimum accuracy of the classification model based on the vis NIR spectra(VNIR–PLSDA)is 62.2%.The model accuracy based on vis–NIR–MIR and OPA fusion spectra is superior to vis–NIR but not as high as MIR classification accuracy in PLSDA models;In RF models,the classification model accuracy of fused spectra is not as good as that of a single spectrum,which does not reflect the advantages of fusion strategies.Because MIR spectral data has already achieved high classification accuracy(MIR–RF overall accuracy of 89.1%),the improvement space of the model is limited.Among the four soil classes,podzols has the best classification accuracy is the RF model based on the MIR spectra,with an accuracy of 97%.Secondly,acrisols has an average classification accuracy of over 86% in the RF model.The luvisols hasan average accuracy in the RF model is 83%,while the overall classification effect is poor for cambisols,which has an average classification accuracy of 69% in the RF model.We found that the accuracy of spectral fusion strategy in spectral classification is not necessarily higher than that of single spectral data classification.The combination of MIR spectral data and RF classification method is a fast,accurate,and convenient strategy for determining soil classes.