Causality analysis and prediction of soil saturated hydraulic conductivity by combining empirical modeling and machine learning techniques

Soil saturated hydraulic conductivity (Ks), as a fundamental property governing the behavior of water in the soil environment, is important to agriculture, environmental quality, water resources management, and engineering practices. However, the intricate interplay between Ks and environmental factors leads to substantial variability, making direct measurement costly and challenging. To provide a cost-effective indirect method with regard to Ks estimation, we proposed a portable modeling framework combing partial least squares structural equation modeling and artificial neural network (PLS-SEM-ANN). The framework was applied to a mountainous watershed characterized by humid climate, steep topography, and diverse land use. This approach demonstrated that the PLS-SEM-ANN was efficient in Ks modeling with a prediction accuracy of more than 80 % (R-2 of 0.989 and 0.862 for model training and validation), which is comparable to that of the conventional artificial neural network (R-2 of 0.993 and 0.843 for model training and validation). According to structural equation modeling, land use significantly affected Ks through direct (beta = 0.237, p < 0.05) and indirect (beta = 0.263, p < 0.05) ways, and soil properties, especially soil particle composition, were the most direct factor affecting Ks (beta = 0.410, p < 0.01), while topography had a lesser effect on Ks (beta = -0.023, p > 0.05). However, Shapley additive explanations (SHAP) analyses of PLS-SEM-ANN revealed significant nonlinear effects of topography on Ks, especially slope (SHAP mean absolute value = 0.92). Moreover, threshold testing showed that there was an abrupt change in the slope-Ks relationship at approximately 21 degrees, with the effect of slope on Ks turning from positive to negative. This study provides an approach with explainable and transparent modeling results for Ks.