Inversion Study Of Soil Salinization In The Yinchuan Plain Based On Spectral Index And Elimination Of Moisture Effects

Soil salinization is an important factor limiting agricultural production safety and ecological environment protection in China.The soil salinity content directly affects soil quality and plant growth and development.Rapid and accurate acquisition of soil salinity is the foundation for soil management and has significant implications for precision agriculture and environmental protection.Hyperspectral remote sensing technology,with its dynamic and efficient characteristics,has been widely used in the inversion research of saline-alkali soils.However,due to the regional heterogeneity of geographical environment,the accuracy of different inversion models varies,and their applicability is somewhat difficult.Therefore,it is essential to explore a high-precision inversion method for soil salinity based on local characteristics to accurately and comprehensively monitor the soil salinization status.This study takes Yinchuan Plain as the research area,using grid-laid soil samples,field and laboratory spectra as data sources,and combines soil salinity measured values.Methods from natural geography,soil science,geographic information science,and mathematical statistics were employed to analyze the statistical characteristics and spectral characteristics of the soil in the study area.The applicability of different hyperspectral data to soil salinity inversion was clarified,and poor hyperspectral information was corrected.The application of fractional-order derivatives combined with two-dimensional spectral indices in soil salinity inversion was explored.A hyperspectral inversion model for soil salinity that removes water interference was established,and spatial distribution prediction was performed using interpolation.This study provides a scientific basis and theoretical support for the accurate and rapid inversion of soil salinity in the Yinchuan Plain.The main results of this study were as follows:(1)The soil salinity in the Yinchuan Plain ranges from 0.30 to 14.98 g·kg-1,and the coefficient of variation was greater than 1,with a strong degree of variation.The degree of salinization was directly proportional to the spectral reflectance.After different spectral transformations,the first-order differential reflectance(FDR)can significantly enhance the sensitivity of spectral information to soil salinity in terms of maximum absolute correlation coefficient(MACC)values and sensitive band response ranges.The two-dimensional correlation of the five spectral indices to soil salinity is ranked as GDVI>RSI>NDSI>DSI>BSI.Compared with the original reflectance(Ref),the response to soil salinity was effectively improved.FDR can explain the gradual process of differential transformation spectral curves from 0 to 1.0 order,and the reflection and absorption characteristics of the spectra were highlighted from 1.0 to 2.0 orders.(2)The inversion accuracy of field spectra for soil salinity was higher than that of laboratory spectra.The partial least square regression(PLSR)and support vector machine(SVM)model validation determination coefficients(Rp2),root mean square errors(RMSE),and relative prediction deviations(RPD)were 0.630,1.44,1.64,and 0.950,0.58,4.09,respectively,while the inversion accuracy of laboratory spectra Rp2 was all less than 0.50.The inversion accuracy of corrected laboratory spectra was generally improved.The segmented correction inversion accuracy of the PLSR model was higher than the global correction,while the global correction inversion accuracy of the SVM model was higher than the segmented correction.The modeling accuracy of the SVM model was generally better than the PLSR model,with the best results for field spectra,followed by laboratory global correction spectra and laboratory segmented correction spectra,and the worst for laboratory spectra.(3)Integer-order differential transformation of field spectra can effectively enhance the spectral information response to soil salinity.With a 0.9-order boundary for fractional-order differentiation,the differential reflectance(DR)was weaker in correlation with soil salinity than the original spectral reflectance when the order was less than 0.9,and it can effectively enhance the correlation with soil salinity when the order was greater than or equal to 0.9.The combination of feature bands selected by fractional-order differentiation combined with two-dimensional spectral indices has a higher correlation with soil salinity than one-dimensional bands.The modeling accuracy of the SVM model was better than the multiple linear regression(MLR)model,with the optimal models for both being 1.1-order differentiation.For the SVM model,Rp2,RMSE,and RPD are 0.839,0.96,and 2.46,respectively,and for the MLR model,Rp2,RMSE,and RPD were 0.730,1.32,and 1.79.(4)Soil moisture content is inversely proportional to spectral reflectance.Spectra have absorption bands near 1,430 nm,1,950 nm,and 2,200 nm,with the main absorption band near 1,950 nm.Longwave drift occurs in each absorption band under different moisture content gradients.Orthogonal signal correction(OSC),FDR,and FDR-OSC can reduce absorption features to varying degrees,with FDR-OSC being the most effective.The correlation of Ref,OSC,FDR,and FDR-OSC with soil moisture content and soil salinity was opposite.The correlation with moisture content was Ref>OSC>FDR>FDR-OSC,while the correlation with soil salinity was FDR-OSC>FDR>OSC>Ref.In the "water removal" mechanism,OSC,FDR,and FDR-OSC reveal more soil salinity-sensitive bands compared to Ref,providing potential for identification.The SVM-based FDR and FDR-OSC "water removal" models for soil salinity have Rc2,Rp2,and RPD values of 0.873,0.950,and 4.09,and 0.952,0.960,and 5.04,respectively,demonstrating strong fitting and predictive capabilities.The FDR-OSC model,FDR model,and measured values of soil salinity were spatially interpolated and predicted using the inverse distance weighting(IDW)method.The interpolation maps of the two models have similar trends to the measured maps,with the FDR-OSC interpolation prediction being more consistent with the measured values,and only a small area of prediction discrepancy.The FDR interpolation prediction was higher for some low values and lower for some high values.