Integration of statistical models and ensemble machine learning algorithms (MLAs) for developing the novel hybrid groundwater potentiality models: a case study of semi-arid watershed in Saudi Arabia

The present study has proposed three novel hybrid models by integrating three traditional ensemble models, such as random forest, logitboost, and naive bayes, and six newly developed ensemble models of rotation forest (RF), such as decision tree (RF-DT), J48 (DF-J48), naive bayes tree (RF-NBT), neural network (RF-NN), M5P (RF-M5P) and REPTree (RF-REPTree), with three statistical models, i.e. weight of evidence, logistic regression and combination of WOE and LR. To predict the groundwater potential, nine groundwater potential conditioning parameters have been created. The Information Gain Ratio has been used to evaluate the impact of each parameter. The ROC curve has been used to validate the models. According to the findings, 15 to 30% of the study area has a very high or high groundwater potentiality. Furthermore, validation results revealed that RF based ensembles models outperformed other standalone models for groundwater potential modelling.

本研究提出三种新型混合模型，将三类传统集成模型——随机森林（Random Forest）、LogitBoost、朴素贝叶斯（Naive Bayes）——以及六种新开发的旋转森林（Rotation Forest，RF）集成模型（即决策树（RF-DT）、J48（DF-J48）、朴素贝叶斯树（RF-NBT）、神经网络（RF-NN）、M5P（RF-M5P）与REPTree（RF-REPTree）），与三种统计模型——证据权重（Weight of Evidence，WOE）、逻辑回归（Logistic Regression，LR）以及WOE与LR的组合——进行集成。为开展地下水潜势预测任务，本研究构建了九项地下水潜势条件参数。采用信息增益比（Information Gain Ratio）评估各参数的影响程度，通过受试者工作特征曲线（Receiver Operating Characteristic Curve，ROC曲线）对模型性能进行验证。研究结果表明，研究区域内15%至30%的区域具备极高或高地下水潜势。此外，验证结果显示，基于旋转森林的集成模型在地下水潜势建模任务中优于其余单一模型。