Groundwater potential mapping using C5.0, random forest, and multivariate adaptive regression spline models in GIS

Area of Science:

• Hydrology and Water Resources Management
• Geospatial Analysis
• Machine Learning Applications

Background:

• Increasing global demand for water necessitates improved understanding of groundwater resources, particularly in arid regions.
• Groundwater is a critical water source, especially in arid climates like Iran.
• Accurate groundwater potential mapping is essential for sustainable water resource management.

Purpose of the Study:

• To generate groundwater potential maps (GPMs) using novel algorithms.
• To validate and compare the performance of C5.0, Random Forest (RF), and Multivariate Adaptive Regression Splines (MARS) for GPM generation.
• To assess the suitability of these machine learning algorithms for groundwater resource assessment in the Mashhad Plain, Iran.

Main Methods:

• A dataset of spring locations and 13 groundwater-conditioning factors (GCFs) was compiled.
• GCFs included altitude, slope, curvature, TWI, land use, lithology, and distances/densities of rivers and faults.
• C5.0, RF, and MARS algorithms were applied to training (70%) and validation (30%) datasets using R statistical software.

Main Results:

• The Multivariate Adaptive Regression Splines (MARS) algorithm achieved the highest performance with an Area Under the Receiver Operating Characteristic Curve (AUC-ROC) of 84.2%.
• Random Forest (RF) and C5.0 algorithms showed acceptable performance with AUC-ROC values of 79.7% and 77.3%, respectively.
• All tested models demonstrated acceptable performance (AUC-ROC > 70%), indicating their utility for GPM generation.

Conclusions:

• The MARS algorithm is highly effective for generating accurate groundwater potential maps.
• The developed methodology and GPMs can be applied to other regions for effective water resource management.
• These GPMs will aid water resource managers in efficient water exploitation and planning.