Mapping Aquifer Vulnerability Indices Using Artificial Intelligence-running Multiple Frameworks (AIMF) with Supervised and Unsupervised Learning

DRASTIC-based vulnerability indices and their variations for an aquifer are investigated in this paper, each of which is regarded as a framework since their rationale of using seven DRASTIC data layers is consensual and lacks empirical or theoretical formulations. The Basic DRASTIC framework (BDF) is implemented by a set of prescribed rules; whereas its three variations involve unsupervised learning from the data, which comprise: (i) learning the rates by the Wilcoxon test (WDF) but using BDF weights; (ii) using BDF rates but learning the weights by Genetic Algorithm (BDF-GA); and (iii) learning rates as in WDF and the weights as in BDF-GA (WDF-GA). These four frameworks are not supervised, but the novelty of the paper is to introduce supervised learning at the second stage by Artificial Intelligence to run Multiple Frameworks (AIMF), for which the paper uses Support Vector Machine (SVM). AIMF uses the outputs of the four frameworks as its input data and a function of observed nitrate-N values as its target data. The AIMF strategy is evaluated in the aquifer of Ardabil plain, which is exposed to anthropogenic contamination such as nitrate-N. The coefficient of correlation (r-values) between the results and nitrate-N values for the above frameworks are: 0.2, 0.37, 0.38 and 0.45; whereas AIMF enhances it to 0.84; attributable to the supervised learning.