关键词: 高光谱, 有机质, 光谱预处理, 变量优选方式, 模型

Abstract: The rapid monitoring of soil organic matter (SOM) content based on hyperspectral data is of great significance for evaluating soil fertility. To achieve rapid and accurate monitoring of SOM in the Yinchuan Plain, we collected 171 soil samples at 0-20 cm in Yinchuan Plain by 5 km×5 km grid method, and measured the SOM content in the laboratory and the spectral reflectance by ASD FieldSpc4 Spectrum Analyzers in the field. After spectral preprocessing and resampling, we obtained five spectral indices, including standard normal variable (SNV), maximum normalization (MAN), first order differential (FDR), logarithm of reciprocal (LR), and reciprocal of logarithm (RL). We determined the sensitive bands of SOM using the correlation coefficient method, and screened out the optimal modeling variables from the sensitive bands by stepwise regression (SR), gray correlation degree (GCD), and principal component analysis (PCA). We established estimation models of SOM by ridge regression (RR), partial squaresregression (PLSR), support vector machine (SVM), and back propagation neural network (BPNN), respectively. The results showed that, compared with resampling spectral reflectance (REF), the correlation between the five spectral indices and SOM did not change significantly after conventional transformation. The spectral index SNV was the common input variable when the optimal modeling variables were extracted by SR, GCD, and PCA. Compared with SR and GCD, the PCA screening approach was the best with respect to the accuracy of model estimation. In the PCASVM models, the model based on spectral index RL had the highest accuracy, with R_c², R_P² and RPD of 0.74, 0.78 and 2.08, respectively. This study determined RL-PCA-SVM as the optimal estimation model by comparing and analyzing the model accuracy under different spectral transformation, variable screening methods and modeling approaches. The model could provide reference for the rapid monitoring of SOM content in Yinchuan Plain and similar areas.

Key words: hyperspectral, organic matter, spectral pretreatment, variables selection method, model.