Ensemble Boosting and Bagging Based Machine Learning Models for Groundwater Potential Prediction
Ensemble Boosting and Bagging Based Machine Learning Models for Groundwater Potential Prediction
复制标题
DOI：
10.1007/s11269-020-02704-3
复制
发表时间：
2020-11-17
影响因子：
4.3
通讯作者：
Rafiei Sardooi, Elham
中科院分区：
文献类型：
--
作者：
Mosavi, Amirhosein;Sajedi Hosseini, Farzaneh;Rafiei Sardooi, Elham
由于对作为主要淡水资源之一的地下水的需求迅速增加，因此迫切需要推进新型预测系统，以更准确地估计地下水潜力，从而进行明智的地下水资源管理。据报道，集成机器学习方法通常可以产生更准确的结果。然而，提出新颖的集成模型以及对这些模型的性能评估进行比较研究对于精确确定合适的方法同样重要。因此，当前的研究旨在提供有关四种集成模型性能的知识，即Boosted广义加性模型（GamBoost）、自适应Boosting分类树（AdaBoost）、Bagged分类和回归树（Bagged CART）以及随机森林（RF）。为了建立模型，使用了 339 个地下水资源的位置和空间地下水潜力调节因子。此后，应用递归特征消除（RFE）方法来识别关键特征。 RFE 指定地下水势建模的最佳特征数量是 15 个变量中的 12 个（平均精度约为 0.84）。建模结果表明，Bagging 模型（即 RF 和 Bagged CART）比 Boosting 模型（即 AdaBoost 和 GamBoost）具有更高的性能。总体而言，RF 模型优于其他模型（准确率 = 0.86，Kappa = 0.67，精度 = 0.85，召回率 = 0.91）。此外，地形位置指数的预测变量、山谷深度、排水密度、海拔和距河流的距离在建模过程中贡献最大。本研究预测的地下水潜力图可以帮助流域和含水层管理领域的水资源管理者和政策制定者保持对这一重要淡水的最佳利用。
Due to the rapidly increasing demand for groundwater, as one of the principal freshwater resources, there is an urge to advance novel prediction systems to more accurately estimate the groundwater potential for an informed groundwater resource management. Ensemble machine learning methods are generally reported to produce more accurate results. However, proposing the novel ensemble models along with running comparative studies for performance evaluation of these models would be equally essential to precisely identify the suitable methods. Thus, the current study is designed to provide knowledge on the performance of the four ensemble models i.e., Boosted generalized additive model (GamBoost), adaptive Boosting classification trees (AdaBoost), Bagged classification and regression trees (Bagged CART), and random forest (RF). To build the models, 339 groundwater resources' locations and the spatial groundwater potential conditioning factors were used. Thereafter, the recursive feature elimination (RFE) method was applied to identify the key features. The RFE specified that the best number of features for groundwater potential modeling was 12 variables among 15 (with a mean Accuracy of about 0.84). The modeling results indicated that the Bagging models (i.e., RF and Bagged CART) had a higher performance than the Boosting models (i.e., AdaBoost and GamBoost). Overall, the RF model outperformed the other models (with accuracy = 0.86, Kappa = 0.67, Precision = 0.85, and Recall = 0.91). Also, the topographic position index's predictive variables, valley depth, drainage density, elevation, and distance from stream had the highest contribution in the modeling process. Groundwater potential maps predicted in this study can help water resources managers and policymakers in the fields of watershed and aquifer management to preserve an optimal exploit from this important freshwater.