Modeling soil water repellency in loess soils of northern Iran using machine learning

Extended Abstract

Introduction

A major hydrological and physical event affecting surface runoff, erosion, and water infiltration is soil water repellency (SWR). Hydrophobic soils reject wetting, therefore causing water droplets to linger on the surface instead of penetrating the soil profile. Particularly in sloping environments and arid ecosystems, this disease causes more overland flow, less water retention, and higher sensitivity to soil loss. SWR is progressively seen in northern Iran, especially in the loess-derived soils of Golestan and Mazandaran provinces, as a result of a mix of climatic conditions and changes in land use. Fine silty to silty-clay textures characterize loess soils in these areas; together with environmental stressors including drought cycles, agricultural development, and deforestation, they aid in the formation and magnification of SWR. Central in the water repellency are organic compounds, especially hydrophobic plant-derived substances like waxes and lignins. Emphasizing the significance of soil chemical composition, many studies have shown that soil organic carbon is strongly positively linked to SWR intensity. Variations in clay content, pH, and electrical conductivity (EC) can also affect SWR patterns. Although SWR is very important in soil degradation processes, little research has been done employing sophisticated data-driven techniques to forecast its spatial variability. Machine learning (ML) algorithms like Decision Tree (DT), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) provide strong means for modeling complicated soil behavior. Using these algorithms, the current study attempts to forecast SWR in loess soils based on a thorough set of physicochemical parameters, therefore helping to justify improved soil management and erosion control measures.

Materials and Methods

Northern Iran served as the location for the study, which looked at specific loess terrains in Golestan and Mazandaran provinces. From many sites including Gorgan, Maraveh Tappeh, Neka, Sari, and Amol, 45 surface soil samples (depth 0–10 cm) were gathered. While minimizing confounding effects, sampling places were chosen to record changes in topography, land use, and vegetation. Key soil physicochemical characteristics assessed were organic carbon (OC), organic matter (OM), electrical conductivity (EC), pH, mean weight diameter (MWD) of soil aggregates, and particle size distribution (sand, silt, clay). With infiltration time recorded up to 3000 seconds, WDPT tests involved dropping 50 μL distilled water droplets on air-dried soil surfaces at room temperature. WDPT values were used as the target variable in model development. Procedures used in laboratories complied with the same criteria used in earlier research. R software was used for data preprocessing. Outlier detection based on interquartile range (IQR), Z-score normalization of numerical variables, and multicollinearity analysis utilizing the Variance Inflation Factor (VIF) were included in this. Categorical variables like soil texture classes were converted to dummy variables using one-hot encoding. Three machine learning approaches—Decision Tree (CART approach), Random Forest (RF), and Extreme Gradient Boosting (XGBoost)—were applied to the dataset, which was randomly separated into 70% training and 30% testing parts. Models were implemented using R packages rpart, randomForest, and xgboost. Through repeating 10-fold cross-validation, hyperparameter tuning was carried out to enhance prediction accuracy.

Results and Discussion

Initial model performance using default settings revealed limited predictive ability across all algorithms. The Decision Tree (DT) model yielded the weakest results with RMSE = 19.55 and R² = 0.02, indicating poor capacity to capture the variability in WDPT values. After hyperparameter optimization, both Random Forest (RF) and XGBoost (XGB) showed significant improvements. The RF model achieved RMSE = 15 and R² = 0.42, while XGB recorded RMSE = 14.7 with the same R², highlighting their comparable predictive power. Feature importance analysis revealed that organic carbon was the most influential predictor of WDPT across all models. Additional influential variables included clay content, sand fraction, EC, OM, and pH, though their relative importance varied by algorithm. In RF, organic matter and sand had high predictive value, whereas in XGB, clay and EC gained prominence. These differences reflect each model's inherent sensitivity to nonlinear interactions. Spatial analysis showed that areas with higher organic carbon content aligned with regions of higher WDPT, confirming the key role of hydrophobic organic compounds in driving soil water repellency. Uncertainty assessment using Bootstrap and Monte Carlo simulations demonstrated that RF was the most stable model, showing the lowest RMSE variability and higher resilience to noisy input data. Overall, the results confirm that machine learning algorithms, especially RF and XGB, can effectively model and interpret the complex interactions influencing soil water repellency in loess landscapes.

Conclusion

This study demonstrated the applicability of advanced machine learning algorithms for modeling soil water repellency (SWR) in loess-derived soils of northern Iran. Among the three tested models, Random Forest (RF) provided the most reliable and stable predictions, with optimal performance metrics (RMSE = 15, R² = 0.42) and low sensitivity to data uncertainty. XGBoost (XGB) also yielded competitive results but showed slightly lower stability under uncertain conditions. The Decision Tree (DT) model, while interpretable, lacked sufficient predictive accuracy for complex, nonlinear relationships. The results confirmed that organic carbon is the dominant driver of SWR in the study area, supporting previous findings regarding the hydrophobic nature of plant-derived organic compounds. Other variables such as clay content, sand fraction, pH, and EC also played important roles depending on the model structure. Differences in variable importance highlight the benefit of using multiple algorithms to obtain a comprehensive understanding of the underlying mechanisms. Uncertainty analysis showed that RF is less susceptible to overfitting and data noise, making it a more robust choice for environmental modeling. Spatial patterns of WDPT and key soil variables revealed strong regional correlations, suggesting the feasibility of using geospatial ML models for site-specific soil management. Future research should explore hybrid models (e.g., RF-XGB) and deep learning architectures (e.g., CNNs or ActionFormer) to enhance predictive power, particularly in dynamic or post-disturbance soil systems. Moreover, integrating multi-temporal datasets could improve the understanding of SWR variability under different environmental and management conditions.