Strategies for Learning Groundwater Potential Modelling Indices under Sparse Data with Supervised and Unsupervised Techniques

Mapping for Groundwater Potential Indices (GPI) is investigated for study areas with sparse data by the customary ten general-purpose data layers with a scoring system of rates and weights but assigning their values give rise to subjectivity. Learning rates/weights from site-specific data reduces subjectivity through unsupervised models. The use of supervised models requires target values, and the paper derives their values from the record at all the productive wells by developing a binary classification model. The paper formulates an Inclusive Multiple Modelling (IMM) strategy to learn from the site data at two levels: at Level 1: two unsupervised ‘base’ models and four supervised ‘base’ models are investigated; at Level 2 the IMM strategies include a supervised ‘combiner’ model, which uses outputs of unsupervised base models; as well as an unsupervised ‘combiner’ model, which uses outputs of supervised base models. Performance metrics are derived by the Receiver Operating Characteristic (ROC) curves and Area Under Curve (AUC). The results show that unsupervised learning at Level 2 (using supervised base models) may reduce subjectivity but even supervised learning at Level 1 can be effective in extracting essential information from target values. Although unsupervised models would extract marginal information from models at Level 1, a supervised model at Level 2 can extract good information from unsupervised models at Level 1.