Estimation of spatiotemporal response of rooted soil using a machine learning approach

基于机器学习算法估算根系土体特性的时空响应

In this study,a machine learning method,i.e.genetic programming(GP),is employed to obtain a simplified statistical model to describe the variation of soil suction in drying cycles using five selected influential parameters.The data used for model development was recorded by an in-situ experiment.The image processing technology is used to quantify several tree canopy parameters.Based on four accuracy metrics,i.e.root mean square error(RMSE),mean absolute percentage error(MAPE),coefficient of determination(R2),and relative error,the performance of the proposed GP model was evaluated.The results indicate that the model can give a reasonable estimation for the spatiotemporal variations of soil suction around a tree with acceptable errors.Global sensitivity analysis for the statistical model obtained using limited data of a specific region demonstrates the drying time as the most influential variable and the initial soil suction as the second most influential variable for the soil suction variations.A case study was conducted using a set of assumed input variable values and validated that the simplified GP model can be used to estimate and predict the spatiotemporal variations of soil suction in rooted soil at a certain range.

目的:在绿色岩土工程中,浅层土体特性通常受到当地气候和覆盖植被的影响。本文旨在探讨自然环境条件下不同植物和大气因素(与树的距离、空气湿度和距离地表的深度等)与土体基质吸力的关系,通过一种机器学习方法建立简化的统计模型,并对浅层根系土体中基质吸力的时空变化进行估算和预测。创新点:1.通过一种机器学习方法(即遗传编程算法)建立土体基质吸力和五个选定的影响因素之间的关系;2.根据建立的统计模型,有效地预测了根系土体内基质吸力的时空变化。方法:1.通过现场监测实验(图3和4),量化土体基质吸力和不同影响参数随时间的变化(图5和6);2.通过机器学习算法,构建土体基质吸力的时空变化与五个选定的影响参数之间的关系,得到一个简化的统计模型(公式(11));3.通过误差分析,验证该简化统计模型在估算和预测土体基质吸力时空变化时的可靠性;4.通过敏感性分析研究不同参数对土体基质吸力时空变化的影响(图9);5.通过案例研究,验证利用该方法对根系土体基质吸力时空变化进行预测的可行性(图11和12)。结论:1.遗传编程算法可以有效地建立土体基质吸力和不同影响参数之间的关系,并能给出相应的数学公式以对土体基质吸力的时空变化进行可靠的估算和预测;2.基于方差的全局敏感性分析方法发现干循环时间和初始基质吸力对土体基质吸力的时空变化有重要影响,而且其他的植物和大气相关参数对土体基质吸力的时空变化也有不可忽视的影响;3.案例研究结果表明,本文所提方法可用于预测土体基质吸力的时空变化。