Variations in Soil Texture Using Machine Learning Models Based on the Spectral Response of Soil Samples in the Visible-Near-Infrared (Vis-NIR) Region

In this study, we systematically assessed the predictive performance of two machine learning models; Random Forest (RF) and Partial Least Squares Regression (PLSR), utilizing various pre-processing techniques for the determination of soil properties, including sand, silt and clay fractions. The evaluation was conducted based on key performance metrics, with the following values recorded: RF exhibited outstanding results with high R2 values (0.88 for calibration and 0.55 for validation) and low RMSE (0.48 for calibration and 1.81 for validation), especially excelling in predicting clay content (R2: 0.92 in calibration and 0.78 in validation). Moreover, the RF model demonstrated impressive RPD (12.5 in calibration and 4.55 in validation) and RPIQ (4.87 in calibration and 4.13 in validation) values for clay. PLSR demonstrated moderate performance, achieving acceptable R2 values for sand, silt and clay fractions, with the highest R2 value of 0.79 achieved in sand content prediction using MSC pre-processing. The RPD and RPIQ scores supported the model's reliability. These findings offer valuable guidance for selecting suitable models and pre-processing techniques for soil property prediction, with Random Forest emerging as the top choice for accurate and reliable results.