Formulating a strategy to combine artificial intelligence models using Bayesian model averaging to study a distressed aquifer with sparse data availability

A modelling strategy is formulated, which collectively consists of separate Multiple Models (MM) and uses Bayesian Model Averaging (BMA) to combine these MMs to learn from data. The procedure is at two levels: at Level 1, three Artificial Intelligence (AI) models are constructed, which comprise Artificial Neural Network (ANN), Sugeno Fuzzy Logic (SFL) and Multiple-Neuro-Fuzzy (Multi-NF) but their outputs are directed to BMA at the next level; at Level 2, BMA is used to combine ANN, SFL and Multi-NF for better predictions and with facilities for quantifying uncertainty. The model performance is tested using the data from Urmia aquifer in the West Azerbaijan province, northwest Iran, where due to the absence of participatory water usage management practices both Lake Urmia and its surrounding 12 aquifers (including the study area) are distressed. The modelling strategy and its results on Urmia aquifer provides an insight to the study area and will be used to investigate ways of arresting the decline in the water table of the aquifer but this should be feasible only by developing a series of management plans, including basin management plans, aquifer management plans, drought plans, water cycle studies. Under such an integrated management system, the model developed here is demonstrably well-placed to serve as an operational management tool for the aquifer.