Comparative study for coastal aquifer vulnerability assessment using deep learning and metaheuristic algorithms.

Area of Science:

• Hydrogeology
• Environmental Science
• Water Resource Management

Background:

• Coastal aquifer vulnerability assessment (CAVA) is critical for managing seawater intrusion (SWI).
• Existing models like the original GALDIT model (OGM) require enhanced accuracy and optimized parameters.
• Deep learning applications in modifying CAVA model weights and rates are underexplored.

Purpose of the Study:

• To investigate the vulnerability of the Lahijan coastal aquifer to SWI using advanced modeling techniques.
• To compare the performance of hybrid optimization models (OGM-BBO, OGM-GWO) with a deep learning approach (CNN).
• To develop an accurate vulnerability map (VM) for effective coastal resource management.

Main Methods:

• Applied the original GALDIT model (OGM) and assessed parameter significance using mean decrease accuracy (MDA).
• Introduced biogeography-based optimization (BBO) and gray wolf optimization (GWO) to create hybrid OGM-BBO and OGM-GWO models.
• Developed a novel convolutional neural network (CNN) algorithm to generate a CNN-based vulnerability map (VM_CNN).

Main Results:

• The CNN-based VM achieved a superior area under the curve (AUC) of 0.982, outperforming OGM-BBO (0.794) and OGM-GWO (0.835).
• The CNN-based VM identified 41% of the aquifer as having very high to high vulnerability to SWI, concentrated near the coastline.
• 32% of the aquifer showed very low to low vulnerability, primarily in southern and southwestern areas.

Conclusions:

• The convolutional neural network model offers a significant advancement in coastal aquifer vulnerability assessment for seawater intrusion.
• The proposed deep learning approach provides a highly accurate tool for identifying at-risk areas, aiding land use planners and policymakers.
• This methodology can be extended to other coastal aquifers, enhancing understanding of aquifer vulnerability and contamination risks.