Multiple-depth modeling of soil organic carbon using visible-near infrared spectroscopy

This paper evaluates the capability of visible-near-infrared (VIS-NIR) spectroscopy to estimate soil organic carbon (SOC) at multiple depths including 0-15, 15-40, 40-60, and 60-80 cm. Four modeling algorithms, namely partial least squares regression (PLSR), principal component regression (PCR), support vector regression (SVR), and random forest (RF) were implemented calibrated to process the spectroscopy data. Overall, 120 soil samples were taken from 30 profiles at the depth of 0-80 cm. We implemented the four models considering different pre-processing techniques including Savitzky-Golay first deviation (SGD), normalization (N), and standard normal variate transformation (SNV). Results revealed that the RF model outperformed other models and the highest accuracy was reached with no pre-processing for all depths excluding 40-60 cm, where the R-2 and RMSE were between 0.55-0.77 and 0.75-0.84% respectively. For the depth of 40-60 cm, the maximum accuracy was observed when SGD pre-processing was applied, resulting in R-2=0.73 and RMSE = 0.78%. Generally, our findings indicate that the spectral data can provide useful information to predict SOC at multiple depths.

Automated summary

This study evaluates the capability of visible-near-infrared spectroscopy to estimate soil organic carbon at different depths. The results show that the random forest model performs better than other models, and the highest accuracy is achieved with no preprocessing for most depths. The study demonstrates that spectral data can provide useful information for predicting soil organic carbon content at different depths.