The absorbance spectra for air-dried and ground soil samples from Ontario, Canada were collected in the visible and near-infrared (VIS-NIR) region from 343 to 2200 nm. The study examined thirteen combination of six preprocessing (1st derivative, 2nd derivative, Savitzky-Golay, Gap, SNV and Detrend) method included in 'prospectr' R package along with four modeling approaches: partial least square regression (PLSR), cubist, random forest (RF), and extreme learning machine (ELM) for prediction of the soil organic matter (SOM). The 1st derivative + gap, 2nd derivative + gap and standard normal variance (SNV) were the best preprocessing algorithms. Thus, only these three preprocessing algorithms along with four modeling approaches were used for prediction of soil pH, electrical conductively (EC), %sand, %silt, %clay, %very coarse sand (VCS), %coarse sand (CS), %medium sand (ms) and %fine sand (fs). The results showed that OM, pH, %sand, %silt and %CS were all predicted with confidence (R2 > 0.60) and the combination of 1st derivative + gap and RF gained the best performance. A detailed comparison of the preprocessing and modeling algorithms for various soil properties in this study demonstrate that for better prediction of soil properties using VIS-NIR spectroscopy requires different preprocessing and modeling algorithms. However, in general RF and 1st derivative + gap can be labeled at the best combination of preprocessing and modelling algorithms.
Keywords: digital soil maps; precision agriculture; proximal soil sensing; soil characterization; soil core profiles.