Prediction of soil salinity with soil-reflected spectra: A comparison of two regression methods

Abstract

To achieve the best high spectral quantitative inversion of salt-affected soils, typical saline-sodic soil was selected from northeast China, and the soil spectra were measured; then, partial least-squares regression (PLSR) models and principle component regression(PCR) models were established for soil spectral reflectance and soil salinity, respectively. Modelling accuracies were compared between two models and conducted with different spectrum processing methods and different sampling intervals. Models based on all of the original spectral bands showed that the PLSR was superior to the PCR; however, after smoothing the spectra data, the PLSR did not continue outperforming the PCR. Models established by various transformed spectra after smoothing did not continue showing superiority of the PCR over the PLSR; therefore, we can conclude that the prediction accuracies of the models were not only determined by the smoothing methods, but also by spectral mathematical transformations. The best model was the PCR based on the median filtering data smoothing technique (MF) + log (1/X) + baseline correction transformation (R2 = 0.7206 and RMSE = 0.3929). To keep the information loss becoming too large, this suggested that an 8 nm sampling interval was the best when using soil spectra to predict soil salinity for both the PLSR and PCR models.