A thorough understanding of soil wetting patterns during infiltration is essential for designing surface drip irrigation systems and placing soil moisture sensors. This study systematically evaluated variations in centroid depth (zc), horizontal (σx) and vertical (σz) standard deviations across nine soil textures, three discharge rates (1, 2, and 3 L·h−1), and three initial soil water contents (30 %, 50 %, 70 % of maximum available water) using Hydrus-2D/3D numerical simulations combined with spatial moment analysis. Based on these results, a machine learning model combining particle swarm optimization (PSO), support vector machine (SVM), and adaptive boosting (AdaBoost) was developed and compared with multiple linear regression (MLR), SVM, and PSO-SVM models. Soil texture and initial water content had greater influence on zc, σx, and σz than discharge rates. The PSO-SVM-AdaBoost model achieved the highest accuracy, with Bias, Root Mean Square Error (RMSE), and the Coefficient of Determination (R2) for the test set of −0.129 cm, 1.139 cm, and 0.989 for zc; −0.034 cm, 0.366 cm, and 0.996 for σx; and −0.169 cm, 1.426 cm, and 0.984 for σz. Furthermore, to address concerns regarding the “black-box” nature of the model, the explainable artificial intelligence (XAI) framework SHapley Additive exPlanations (SHAP) was applied, revealing that cumulative infiltration flux (Q3D) contributed most significantly to zc, σx, and σz, while discharge rates contributed the least. The PSO-SVM-AdaBoost model and its interpretability framework proposed in this study provide technical support for the design of surface drip irrigation systems and the optimal placement of soil moisture sensors.
扫码关注我们
求助内容:
应助结果提醒方式:
