Quantitative Inversion Of Soil Organic Matter In Yinchuan Plain Area Of Ningxia Based On Hyperspectral Remote Sensing

Soil organic matter(SOM),as the main indicator for evaluating soil fertility,has a direct impact on the water and fertilizer retention capacity of soil and the growth of crops.Rapid and accurate acquisition of the distribution status of soil organic matter is of great significance for the precise development of regional agriculture and the rational utilization of land resources.Currently,hyperspectral techniques have been widely used as a rapid and nondestructive monitoring tool for quantitative inversion studies in areas with high soil organic matter content.However,when the SOM content is lower than 20 g·kg-1,the screening of spectral sensitive indicators is difficult and the inversion accuracy of soil organic matter is poor.Therefore,it is necessary to investigate the best spectral sensitivity characteristics and quantitative inversion models based on hyperspectral information at SOM contents below 20 g·kg-1.In this study,171 soil samples were collected using the 5 km×5 km grid method in the Yinchuan Plain as the study area,and the SOM content measured indoors and the 350-2500 hyperspectral reflectance measured in the field were used as the base data sources.To analyze the statistical characteristics of SOM indicators in the region using soil science,geoinformation technology and mathematical statistics,and to pre-process the raw spectral data based on five conventional variations and fractional order differentiation to reveal the spectral response characteristics of SOM;six linear and nonlinear regression models were developed using different variable screening methods,and the best inverse model for soil organic matter based on different spectral parameters was determined after comparative analysis to provide a scientific basis for rapid and accurate estimation in areas with lower SOM content.The main results of this study are summarized as follows:(1)In the study area,SOM content values ranged from 0.75-19.31 g·kg-1,with low organic matter content and moderate variation,and there was inverse relationship between spectral reflectance and SOM content;compared with the four two-dimensional spectral indices correlations established based on the original spectral reflectance,the one-dimensional sensitive band correlation coefficients based on five conventional variations did not show significant improvement;the spectral data were differentiated by 0.20 order,and the data space was expanded,and the spectral interval data values showed an overall tendency of increasing first and then stabilizing constant.(2)The spectral index SNV(standard normal variable)was the common input variable when the optimal modeling variables were extracted by SR(Stepwise regression),GCD(Grey correlation degree)and PCA(Principal component analysis).Compared with SR and GCD,the PCA screening approach achieved the best accuracy of model estimation;by Compared the model accuracy under different spectral transformation and modeling approaches:the PCA-SVM models that based on spectral index RL had the highest accuracy with Rc2,RP2 and RPD of 0.74,0.78 and 2.08,respectively,the model had good quantification capability for SOM.(3)The maximum absolute correlation coefficient(MACC)values between DI(Difference index),RI(Ratio index),NDVI(Normalized difference index),RDVI(Renormalized difference vegetation index),BI(Brightness index),GDVI(Generalized difference vegetation index)and SOM contents showed the tendency of increasing first and then decreasing,with the highest MACC at 1.0,1.2 and 1.6 orders,respectively.Except DI,RI,BI,GDVI and NDVI,based on the 0.2-2.0 order RDVI under fractional order differential variation could be used for subsequent model construction,in which the sensitive bands of MACC values were mainly focused on 400～600 nm and 1 300～1 700 nm.Compared the accuracy of different models based on the same spectral index,the nonlinear model has the best accuracy,in which the RDVI-SVM model has the highest estimation accuracy with RC2 of 0.86,RP2 of 0.87,and RPD of 2.33(>2.0),respectively,the model has excellent quantification capability for SOM.(4)Based on the algorithm optimization that a single spectral index under fractional order differentiation,the MACC values of the optimization indices DI/RDVI,RDVI/NDVI and NDVI/RDVI were 0.9986,0.7621 and 0.7993 generally higher than 0.70,respectively,in which the enhancement of DI under fractional order differentiation joint was significantly better than that of NDVI.The SVM models constructed by the three optimization indices,R2 was higher than 0.80 overall and RPD was higher than 2.00,and the DI/RDVI-SVM model had the best model fitting accuracy at order 0.2 and 0.4,with both R2 reached 0.98,and RPD was4.31 and 4.27,respectively,which could be used for accurate estimation of SOM content in Yinchuan Plain.(5)Compared the SOM prediction models that built based on sensitive bands,spectral indices and optimization indices,the optimization index DI/RDVI-SVM model had the best prediction accuracy at order 0.4,with R2,RC2 and RPD of 0.99,0.99 and 4.27,respectively,which can be used as a powerful tool for the validation of high-altitude remote sensing inversion models in later studies and provide a basis for the large-scale monitoring of regional lower SOM.