Prediction Of Soil Free Iron And Calcium Carbonate Content In Shaanxi Province Based On Hyperspectral Data

Free iron is an important mineral cement in soil,which can reflect soil formation process,environment and climate change.Soil calcium carbonate is significantly influcing numerous soil physical,chemical and biological properties such as soil pH,soil colloid and nutrient available.In this research,two reasons responsible for soil free iron and calcium carbonate were selected as the target properties.Firstly,both of them have important soil science significance.Secondly,few studies have focused on the prediction of free iron and calcium carbonate based on reflectance spectra,and the precision accuracy is not high.The traditional method for detecting free iron and calcium carbonate in soil is time-consuming and laborious.Hyperspectral remote sensing has the advantages of multiple bands and high spectral resolution,which containing rich informations concerning soil physical and chemical properties.This has provided possibility to obtain soil properties quickly,accurately and quantitatively,and it raised new ideas and methods on predication of soil free iron and calcium carbonate based on the hyperspectral data..In this study,a total of 231 soil samples from 10 major soil types in Shaanxi Province were used as research objects.The Reflectance spectrum curve of soil samples were obtained by hyperspectral imager.Differential four mathematical transformations,such as The first-order differential,second-order differential,continuum removal and the first-order logarithm of the original spectral curve were performed.The feature bands were selected by Analysis of Correlation(CA)and Successive Projections Algorithm(SPA),Foul models of partial least squares regression(PLSR),multiple stepwise linear regression(MLSR),random forest(RF)and support vector machine(SVM)were used to construct prediction models of soil free iron and calcium carbonate.The results are as follows:(1)The spectral curves of 10 soil types in Shaanxi Province are roughly the same.There are obvious absorption valleys in the soil spectral curves at 1400 nm,1900nm and 2200 nm.The spectral reflectance of different soil samples is different.The spectral reflectance difference of different soil free iron content in the range of 750-1300 nm is the largest compared with other bands.As the free iron content increases,the spectral reflectance decreases,showing a negative correlation.The correlation between soil calcium carbonate content and spectral reflectance in the range of 1900-2500 nm is larger than that of other bands.With the increase of calcium carbonate content,the spectral reflectance of soil also increases,both of them show a positive correlation.(2)Based on the feature bands selected by CA and SPA,and the spectral data of four mathematical transformation forms,the prediction models of free iron and calcium carbonate were established by using the four modeling methods of PLSR,MLSR,RFR and SVR.The second-order differential spectral transformation has the highest precision in the soil free iron prediction model.The highest prediction accuracy in the calcium carbonate prediction model is the reciprocal first-order differential spectral transformation;the predictions of the other three models based on the SPA method are simulated except the random forest model.The results are better than the CA method;comprehensively comparing the four models,the SPA-PLS model is used to model the soil free iron and calcium carbonate based on their respective optimal spectral transformation methods with the highest accuracy,which is the best model in all models.The prediction model of soil free iron R2 and RPD are 0.89 and2.97,respectively.The calcium carbonate prediction models R2 and RPD are 0.90 and 3.01,respectively.(3)The classification method based on spectral angle matching-spectral correlation coefficient measurement after spectral feature parameter extraction divides the spectrum into three categories,based on the classified spectral data to predict soil calcium carbonate content,three types of RFR model modeling set and verification set The determination coefficient R2 is about 0.90,and the RPD values are all over 2.0,which is much better than the unclassified global modeling accuracy.It proves the applicability of spectral angle matching-spectral correlation coefficient measure in soil spectral classification modeling.