基于机器学习算法的非饱和土水特征曲线预测

Prediction of Unsaturated Soil Water Characteristic Curve Based on Machine Learning Algorithms

摘要:

土水特征曲线(SWCC)是研究非饱和土渗透、强度预测与本构关系的基础。机器学习算法具有高效处理大量数据和特征提取等特点。本文采用六种机器学习算法(四种集成学习和两种传统机器学习算法)对美国非饱和土数据库中的154条SWCC包含1976个数据点进行模拟；并使用四个性能评价指标(R2、EVS、MAE和RMSE)评价算法的性能。选取两种数据输入的方式：对压力水头进行对数处理和未处理两类。结果表明，在两种输入情况下，对LightGBM、XGB、RF和AdaBoost算法的影响很小；但是对GPR和SVM两种传统机器学习算法的影响很大，在未进行对数处理情况下，R2降低明显甚至会出现无法模拟SWCC的情况。此外，LightGBM对SWCC测试集的模拟效果上均优于其他模型，拥有高的趋势评价指标(R2和EVS)和低的误差测量指标(MAE和RMSE)；六种算法对SWCC模拟的优劣的排列顺序依次为：LightGBM、GPR、XGB、RF、AdaBoost和SVM。最后，利用已训练好的LightGBM模型对9条不包含在数据库内的SWCC数据进行预测，结果显示LightGBM能够较好地预测非饱和土的土水特性。研究结果对提升不同类型土的SWCC预测具有重要的指导意义。

Abstract:

The soil water characteristic curve (SWCC) is the foundation for studying the permeability, strength prediction, and constitutive relationships of unsaturated soils. The machine learning algorithm has the characteristics of efficient processing of large amounts of data and feature extraction. Six machine learning algorithms (four ensemble learning algorithms and two traditional machine learning algorithms) were utilized to model 154 SWCCs comprising 1976 data points sourced from the American Unsaturated Soil Database. The performance of the algorithms was assessed using four performance evaluation indicators (R2, EVS, MAE and RMSE). Two types of data input methods were selected: logarithmic processing of pressure head and untreated. The results indicate that, under the two input types, the impact on the LightGBM, GPR, XGB and AdaBoost algorithms is minimal; however, in the case where pressure head is not logarithmically processed, the impact on the GPR and SVM two traditional machine learning algorithms is significant, R2 drops sharply and it may even result in the inability to model SWCC. Additionally, LightGBM outperforms other models in simulating the SWCC test set, with high trend evaluation indicators (R2 and EVS) and low error measurement indicators (MAE and RMSE). The ranking of the six algorithms in terms of the quality of SWCC simulation is as follows: LightGBM, GPR, XGB, RF, AdaBoost and SVM. Finally, utilizing the LightGBM model trained on the aforementioned database, predictions were made for 9 SWCCs not included in the database. The study revealed that LightGBM can effectively predict the soil water characteristics of unsaturated soils. These research findings have important implications for improving SWCC models for different types of soils.