Objective  A hyperspectral prediction model for organic matter content of soil developed from loess-like parent material in Liaoning Province was established for rapid acquiring contents of soil organic matter (SOM).

  Method  Samples were collected from soils developed from loess-like parent material, and their SOM contents and hyperspectral data were determined. The original spectra and its six spectral transformations of first-order differential, second-order differential, inverse logarithmic, inverse logarithmic first-order differential and inverse logarithmic second-order differential were selected as independent variables, to conduct correlation analysis with SOM content. The characteristic bands in the spectra data were selected, and three linear models for hyperspectral prediction of SOM content were developed by using multiple stepwise linear regression (SMLR), partial least squares regression (PLSR) and principal component regression (PCR), respectively. While nonlinear model fitting by support vector machine (SVM) was also performed.

  Results  The SOM content was negatively correlated with spectral reflectance. The different mathematical treatments of the spectra could improve the correlation between SOM content and spectral reflectance, especial for the first-order differential and second-order differential treatments. the model accuracy of the same spectral in different models differed significantly, the PLSR model with the first-order differentiation of the original spectral reflectance as the independent variable had the highest accuracy, and the coefficients of determination (R²) of the modeling set and validation set were 0.958 and 0.976. The test accuracies of the best prediction models established by the three linear methods were: PLSR > SMLR > PCR.

  Conclusion  The PLSR model was the optimal model for predicting the organic matter content of soil developed from loess-like parent material in Liaoning Province, and the model based on the characteristic bands was better than that based on full bands. The prediction accuracy of SVM nonlinear model was lower.