Explainable Transfer Learning Framework for Water Quality Prediction in Data-Scarce Developing Regions

Safe drinking water access continues to be a preva-lent public health challenge for developing nations, in which continuous water quality monitoring becomes infeasible due to the lack of resources, laboratory facilities, and sensors. Machine learning methods have been successfully applied for predicting water quality metrics such as dissolved oxygen (DO), turbidity, and pH. Nonetheless, traditional supervised machine learning models require sufficient amounts of labeled data (greater than 1000 instances) for optimal generalization. Due to the data-scarcity problem, only 100–200 historical instances are usually available for training. As a result, overfitting becomes inevitable. This paper proposes an Explainable Transfer Learning (XTL) approach for water quality prediction called XTL-WQ. We intro-duce three techniques: (i) a pre-trained Long Short-Term Mem-ory (LSTM) model; (ii) Maximum Mean Discrepancy (MMD) for domain adaptation; and (iii) SHapley Additive exPlanations (SHAP) for interpretability. On DO prediction, we achieve an RMSE of 0.42 mg/L with XTL-WQ, compared to 0.61 mg/L for LSTM baseline (+31% improvement). Under extreme data scarcity (less than 30 instances), our model achieves an RMSE of 0.51, while baselines exceed 0.75 mg/L.