In arid and semi-arid regions, the limited availability of surface water resources makes groundwater a critical source for domestic, agricultural, and industrial use. However, in Djibouti, unplanned and excessive exploitation of groundwater has led to aquifer depletion and seawater intrusion. Therefore, accurate and reliable prediction of groundwater potential represents an urgent need for sustainable water management. In this study, a comprehensive dataset comprising 14 hydrogeological, topographic, climatic, and remote sensing-based variables was developed to model groundwater potential. To predict groundwater potential, the Stochastic Gradient Boosting (SGB) algorithm was applied, while hyperparameter optimization was carried out using three different metaheuristic approaches, namely Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and Gravitational Search Algorithm (GSA). Model performance was evaluated using accuracy, precision, recall, F1-score, and ROC/AUC metrics. Results demonstrated that the GSA-SGB model exhibited the most balanced performance on the test set, with 78.6% precision, 73.6% recall, and an F1-score of 0.749. Additionally, the PSO-SGB model achieved the highest classification accuracy (0.8621), while runtime analysis revealed that PSO was the fastest (195 s) and GA the slowest (627 s). ROC analysis indicated that the highest AUC value (0.813) belonged to the GSA-SGB model. SHapley Additive exPlanations analysis further revealed that Normalized Difference Vegetation Index and the sediment transport index were the most influential factors, with low slope and high precipitation conditions significantly enhancing groundwater potential. Overall, the findings highlight that the integration of machine learning with metaheuristic optimization techniques provides a powerful decision-support framework for predicting groundwater potential in Djibouti.