Hyperspectral estimation of soil organic matter and clay content in loess plateau of China

Visible and near-infrared reflectance (Vis-NIR) spectroscopy is considered a promising tool for the estimation of soil properties. Soil clay content and soil organic matter (SOM) are main components affecting soil spectra. Accurate assessment of clay content and SOM is essential before achieving accurate prediction for other soil properties. Selecting the proper spectral transformation technique and optimal calibration method are important processes to improve model performance. In this study, a total of 240 soil samples were collected from the main area of winter wheat (Triticum aestivum L.) fields in the Southwest region of Shanxi province, northern China. Six spectral pre-treatments and three multivariate methods were utilized to realize the estimation of clay content and SOM. Finally, the important spectral wavelengths were identified as 440, 762, 1,150, 1,410, 1,460, 1,860, 1,900, 2,250, 2,400 nm for clay content and 410, 450, 550, 625, 780, 850, 1,410, 1,670, 1,730, 1,860, 1,910, 1,960, 2,250 nm for SOM. Specifically, the wavelengths around 440 (450), 1,900 (1,910) nm and wavebands of 1,410, 1,860, and 2,250 nm were highly related to both clay content and SOM. The optimal prediction was obtained when multiple linear regression (MLR) was combined with standard normal variate (SNV) pre-processing (R-2 = .714, RMSE = 3.982, RPD = 1.584) for clay content and multiplicative scatter correction (MSC) pre-processing (R-2 = .856, RMSE = 2.994, RPD = 2.443) for SOM. This study implied that spectral transformation had an evident effect on spectral curves shape, correlation, and model performance. The choice of pre-processing transformation should depend on the multivariate technique which has a determined ability to improve the model accuracy.