A mixture learning strategy for predicting aquifer permeability coefficient K

Abstract

Aquifers permeability coefficient (K) is critical for understanding, managing, and protecting groundwater resources. However, obtaining reliable K values directly from pumping tests is costly and time-consuming, often yielding suboptimal results that lead to significant financial losses. Recent advances in machine learning offer an alternative, cost-effective approach for estimating K. Yet, the primary challenge lies in the substantial proportion of missing K data, as K measurements can only be recorded in aquifer layers. Such sparse and incomplete data severely limit the effectiveness of classical supervised learning methods. To address this challenge, we propose a mixture learning strategy (MXS) that combines unsupervised and supervised techniques to improve K prediction. First, a K-Means clustering approach is applied to delineate a naïve group of aquifers (NGA), effectively generating proxy labels for layers where direct K measurements are unavailable. Next, these NGA labels are integrated with existing K values to form enhanced input features for supervised prediction. We then apply support vector machines (SVMs) and extreme gradient boosting (XGB) to predict K more accurately. Experimental results show that both SVMs and XGB achieve prediction accuracies exceeding 80% when evaluated using confusion matrices and micro- and macro-averaged precision-recall metrics. Testing the MXS approach on an independent borehole dataset confirms its robustness and effectiveness. By enabling accurate K predictions in the presence of significant data gaps, MXS supports more informed decision-making, reduces the likelihood of unsuccessful pumping tests, and aids in the sustainable planning and management of groundwater resources.