Formulating Convolutional Neural Network for mapping total aquifer vulnerability to pollution.

Area of Science:

• Environmental Science
• Hydrogeology
• Artificial Intelligence

Background:

• Aquifer vulnerability mapping is crucial for pollution control.
• Traditional methods like the basic DRASTIC framework (BDF) are subjective.
• Developing objective and accurate vulnerability assessment tools is essential.

Purpose of the Study:

• To formulate an AI-based strategy using Convolutional Neural Networks (CNNs) to reduce subjectivity in aquifer vulnerability mapping.
• To develop and test three CNN models for calculating intrinsic (IVI), specific (SVI), and total vulnerability indices (TVI).
• To compare the performance of CNN-derived indices against the BDF using nitrate concentrations.

Main Methods:

• Formulation of an AI modeling strategy using CNNs.
• Training and testing three distinct CNN models for IVI, SVI, and TVI.
• Application of the models to an unconfined aquifer in northwest Iran.
• Evaluation of vulnerability indices using receiver operating characteristic curves and area under the curve (AUC) against nitrate concentrations.

Main Results:

• CNN models demonstrated significantly improved performance over BDF, with AUC values of 0.91 (IVI), 0.95 (SVI), and 0.95 (TVI) compared to 0.81 for BDF.
• While IVI, SVI, and TVI showed similar predictive performance (AUC), their spatial patterns differed.
• Hotspots of aquifer vulnerability were delineated within the study area.

Conclusions:

• CNN-based approaches offer a substantial improvement over traditional methods for aquifer vulnerability mapping, reducing subjectivity.
• The developed CNN models provide accurate and objective assessments of intrinsic, specific, and total vulnerability indices.
• Despite comparable AUC values, the distinct spatial patterns of IVI, SVI, and TVI highlight the importance of considering different vulnerability definitions for targeted management strategies.