基于多种高光谱指标反演冻结土壤含水率的研究

Inverting Frozen Soil Moisture Content Based on Various Hyperspectral Indexes

在中国季节性冻融区如何高效快速地监测冻结土壤的含水率至关重要。以内蒙古河套灌区典型灌域土壤为对象,在实验室分梯度配置不同含水率的土样,通过地物光谱仪测定土壤冻结与未冻结状态下的高光谱数据,处理得到原始光谱反射率(Raw Spectral Reflectance,REF)、一阶微分反射率(First-order Differential Reflectance,FDR)和二阶微分反射率(Second-order Differential Reflectance,SDR)、标准正态变量变换(Standard Normal Variable Reflectance,SNV)和倒数之对数变换(Logarithm of Reciprocal,LR)5种光谱指标,采用偏最小二乘回归法(Partial Least Squares Regression,PLSR)、多元逐步回归法(Multiple Stepwise Regression,MSR)、支持向量机法(Support Vecor Machine,SVM)和一元线性回归法(Unary Linear Regression,ULR),构建基于不同光谱指标的土壤含水率高光谱反演模型并进行回归分析。结果表明,冻结状态下基于REF、FDR和SDR指标构建的模型反演精度低于未冻结状态,REF的验证集决定系数(Rp^(2))的最大值为未冻结状态下PLSR的0.952,最小值为冻结状态下ULR的0.621;FDR的Rp^(2)最大值为未冻结状态下SVM的0.955,最小值为冻结状态下MSR的0.618;SDR的Rp^(2)最大值为未冻结状态下的SVM的0.858,最小值为未冻结状态下PLSR的0.252。未冻结状态下SNV、LR指标的反演精度略低于冻结状态,SNV的Rp^(2)最大值为冻结状态下PLSR的0.796,最小值为未冻结状态下ULR的0.621;LR的Rp^(2)最大值为冻结状态下MSR的0.789,最小值为未冻结状态下ULR的0.667;未冻结状态下最佳模型组合为FDR-SVM,Rp^(2)为0.955,冻结状态下最佳组合模型为REF-PLSR,Rp^(2)为0.799。研究成果可为土壤冻结状态下利用高光谱遥感技术监测土壤含水率提供一定的技术支撑。

In the seasonal freeze-thaw area of China,how to efficiently and quickly monitor the soil moisture content of frozen soil is very important.In this study,soil samples were taken from the Shahaoqu irrigation area of Hetao Irrigation District in Inner Mongolia.Different soil moisture contents were graded in the laboratory,and the experimental hyperspectral data were obtained by ASD Field Spec 3 spectrometer under frozen and unfrozen conditions.After a series of pre-processing,the original spectral reflectance(REF),first-order differential reflectance(FDR)and second-order differential reflectance(SDR),standard normal variables(SNV)and logarithmic transformation of inverse(LR)were used to construct the soil water content inversion models,including Partial Least Squares Regression(PLSR),Multiple Stepwise Regression(MSR),Support Vecor Machine(SVM),and Unary Linear Regression(ULR).The results showed that the inversion accuracy of the model based on REF,FDR and SDR in the frozen state was lower than that in the unfrozen state.The maximum value of the validation set determination coefficient(Rp^(2))of REF was 0.952 of PLSR in the unfrozen state,and the minimum value was 0.621 of ULR in the frozen state.The maximum Rp^(2)value of FDR was 0.955 of SVM in unfrozen state,and the minimum Rp^(2)value was 0.618 of MSR in frozen state.The maximum Rp^(2)value of SDR was 0.858 of SVM in unfrozen state,and the minimum Rp^(2)value was 0.252 of PLSR in unfrozen state.The inversion accuracy of SNV and LR indexes in unfrozen state was slightly lower than that in frozen state.The maximum Rp^(2)of SNV was 0.796 of PLSR in frozen state,and the minimum Rp^(2)of SNV was 0.621 of ULR in unfrozen state.The maximum Rp^(2)of LR was 0.789 of MSR in frozen state,and the minimum Rp^(2)of LR was 0.667 of ULR in unfrozen state.The optimal model combination in the unfrozen state was FDR-SVM with Rp^(2)of 0.955,and the optimal model combination in the frozen state was REF-PLSR with Rp^(2)of 0.799.The research results can provide some technical support for monitoring soil moisture content by hyperspectral remote sensing technology in frozen soil.