Using algorithmic game theory to improve supervised machine learning: A novel applicability approach in flood...

Area of Science:

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

Background:

• Flood susceptibility mapping is crucial for effective water resource management and disaster preparedness.
• Integrating Game Theory (GT) with Machine Learning Algorithms (MLA) offers novel approaches to enhance flood prediction accuracy.

Purpose of the Study:

• To apply Condorcet and Borda scoring algorithms (GT) for flood point determination and Flood Susceptibility Mapping (FSM).
• To evaluate the performance of MLA including Random Forest (RF), Support Vector Regression (SVR), Support Vector Machine (SVM), and K-Nearest Neighbors (KNN) in FSM.
• To identify optimal hybrid MLA-GT models for accurate flood susceptibility assessment in the Cheshmeh-Kileh watershed.

Main Methods:

• Quantification of 14 flood susceptibility conditioning factors (e.g., NDVI, Elevation, Slope, Precipitation).
• Identification of flood and non-flood points using Condorcet and Borda GT algorithms.
• Distributional mapping of flood susceptibility using RF, SVR, SVM, and KNN MLAs.
• Development and evaluation of hybrid MLA-GT models (e.g., RF-Condorcet, RF-Borda).

Main Results:

• The RF-Condorcet and RF-Borda models were identified as the most optimal hybrid MLA-GT approaches for FSM.
• Upstream sub-watersheds were found to be highly susceptible to flooding.
• NDVI and forest cover were the most effective conditioning factors across different classification methods.
• The Condorcet algorithm demonstrated higher classification accuracy compared to the Borda scoring algorithm.

Conclusions:

• Hybrid models combining MLA and GT, particularly RF-Condorcet, offer superior performance for flood susceptibility mapping.
• Effective flood risk management requires considering factors like NDVI and forest cover, especially in upstream areas.
• The Condorcet algorithm provides a more reliable basis for flood susceptibility classification than the Borda algorithm.