Spatial Mapping of Groundwater Potential (GWP) Identification Using Random Forest Machine Learning in Kedah, Malaysia

Abstract

Groundwater is a critical source of freshwater in Malaysia, where rapid urban growth, agricultural expansion, and recurrent shortages of surface water have increased relianced on aquifers. Effective groundwater potential (GWP) mapping is therefore essential to identify suitable zones for sustainable extraction and water security planning. This study aims to determine the GWP area in Kedah by using random forest (RF) machine learning techniques involving 15 groundwater conditioning parameters covering topography, hydrogeology, and environmental factors. A total of 350 tube well locations was partitioned into 70:30 for training and testing dataset. The identified GWP area was classified into five different classes: very high, high, medium, low, and very low. It was found about 30.97% covering an area of 2,798.18 km² is considered as the highest groundwater area in the western and central part of Kedah. Meanwhile, the lowest groundwater area which is about 25.25% is in the northeast part of Kedah covering an area of 2,281.18 km². The performance of the RF model was validated using several evaluation metrics for both training and testing dataset: accuracy, precision, sensitivity, specificity, F1-score and kappa. The validation using receiver operating characteristics (ROC) demonstrated strong discriminative ability with the area under curve (AUC) values of 0.95 (training) and 0.90 (testing). Feature importance analysis revealed elevation was found to be the highest influencing contributors, while lithology was found to be the least influencing contributors for the model’s performances. Overall, the findings highlight the effectiveness of integrating geospatial and machine learning techniques in GWP studies that provides a robust framework which contributes significantly towards sustainable groundwater management strategies.