Hyperspectral Inversion Of Inorganic Carbon In Salinized Soil In Southern Xinjiang Desert

Soil inorganic carbon(SIC)is a key index to maintain the carbon cycle in fragile ecosystems in arid and semi-arid regions,which is related to the balance of carbon cycle and carbon sequestration rate in terrestrial ecosystems.Since the industrial Revolution,people's production and living behaviors have discharged a large amount of greenhouse gases to the atmosphere and accelerated the global warming rate,which is not conducive to the sustainable development of the environment,and has a great negative impact on the absorption of atmospheric carbon dioxide by soil inorganic carbon to maintain the carbon cycle stability of terrestrial ecosystems.Therefore,in the context of global warming,rapid detection of soil inorganic carbon content in arid and semi-arid areas is of great significance for the development of precision agriculture,maintenance of fragile ecosystems and environmental stability in arid areas of China.Traditional laboratory methods to determine soil inorganic carbon content are time-consuming,laborious and environmentally unfriendly.With the development of remote sensing technology and computer technology,hyperspectral technology and machine learning algorithm have been vigorously promoted in the study of soil spectrum and soil physical and chemical properties.This study investigated four data after pretreatment of soil spectral reflectance data and on the basis of using two spectral dimension reduction method for dimension reduction band respectively set up four kinds of soil inorganic carbon quantitative estimates of the model,a total of 48 models for predicting the soil inorganic carbon,choose the most accurate model as the soil inorganic carbon inversion model.The main research results are as follows:(1)Considering the obvious noise in the original spectrum,the spectral reflectance of 350-399 nm and2401-2500 nm is removed,and the spectral reflectance data of 400-2400 nm wavelength is retained.Multivariate scattering correction,Savitzky-Golay smoothing,first-order differentiation and maximum and minimum normalization pretreatment can effectively reduce spectral signal noise,which is conducive to the subsequent modeling of soil inorganic carbon content.There was a significant correlation between hyperspectral data and soil inorganic carbon content after different spectral data pretreatment methods.(2)The continuous projection algorithm can extract 4,8,3 and 8 sensitive bands respectively from the soil spectral data after multiple scattering correction,Savitzky-Golay smoothing,first-order differentiation and maximum and minimum normalization,indicating that the continuous projection algorithm has the ability to effectively reduce the spectral dimension.The ratio of variance of the first three principal components to variance and of the ten principal components screened by principal component analysis were all higher than 0.85,indicating that principal component analysis can effectively reduce the dimension of soil inorganic carbon spectrum.(3)After four pretreatment methods,the sensitive bands screened by the continuous projection algorithm and the ten principal component bands screened by the principal component analysis method,partial least squares regression,BP neural network,random forest and support vector machine models were established for comparative analysis of spectral reflectance.In the random forest model,the first-order differential model of soil spectral reflectance has the highest accuracy,with R2 and RMSE of 0.73 and 0.22 respectively.After first-order differential,soil inorganic carbon reflectance has the ability to predict soil inorganic carbon content combined with the random forest inversion model.Therefore,first-order differential and stochastic forest model can be used to establish the inversion model of soil inorganic carbon content.(4)The R2,RMSE and RPD of the inversion model validation set of soil inorganic carbon content constructed by combining random forest with first-order differential treatment were 0.82,2.76 and 1.46.This indicates that there is a relatively good linear relationship between the predicted value and the measured value,that is,the accuracy of the inversion model is reliable and the estimation ability of the model is stable,which can fully meet the requirements of the inversion modeling of soil inorganic carbon content in southern Xinjiang desert area.Compared with the BP neural network and support vector machine regression model selected in this paper,it has the characteristics of high precision,easy to adjust parameters and strong robustness of the model.(5)A soil inorganic carbon inversion model with high accuracy and moderate estimation ability can be established by combining the processed spectral reflectance with machine learning modeling method.The establishment of machine learning model has the ability to estimate soil inorganic carbon content,which plays an important role in soil carbon cycle,and provides a scientific reference for the actual interception of atmospheric carbon dioxide,and provides a new idea and method for large-scale and efficient estimation of soil physical and chemical properties.In this paper,a series of soil pretreatment methods and dimensionality reduction methods were combined with partial least squares regression model and machine learning model to obtain an appropriate quantitative inversion modeling process of soil inorganic carbon in the desert area through spectral measurement and laboratory measurement of soil samples collected in the field.An inversion model of soil inorganic carbon based on first-order differential and random forest model with high precision and stability was obtained.This is of great significance for maintaining the stability of carbon sequestration in desert ecosystems and exploring the mechanism of soil inorganic carbon change in desert soils.