A novel framework for addressing uncertainties in machine learning-based geospatial approaches for flood prediction.

Area of Science:

• Environmental Science
• Geospatial Analysis
• Machine Learning Applications

Background:

• Machine learning (ML) models are increasingly used for flood susceptibility modeling (FSM), but often produce uncertain spatial predictions.
• Addressing spatial disagreement among different FSMs is complex, hindering reliable flood risk assessment.
• Existing FSMs, while accurate, show significant spatial discrepancies in flood projection outcomes.

Purpose of the Study:

• To present a novel framework for reducing spatial disagreement among multiple ML-based FSMs.
• To develop an optimized hybrid model by integrating the outputs of four distinct FSMs.
• To enhance the reliability and spatial consistency of flood susceptibility predictions.

Main Methods:

• Four ML-based FSMs were employed: Random Forest (RF), K-Nearest Neighbor (KNN), Multilayer Perceptron (MLP), and a hybridized Genetic Algorithm-Gaussian Radial Basis Function-Support Vector Regression (GA-RBF-SVR).
• A case study was conducted in the southwest coastal region of Bangladesh.
• An optimized model was created by combining the predictions of the four individual FSMs to improve spatial agreement.

Main Results:

• All ML-based models predicted a similar flood potential area, covering approximately 60% of the total land.
• Despite high prediction accuracy, significant spatial discrepancies were observed between models (correlation coefficients from 0.62 to 0.91).
• The optimized model demonstrated enhanced prediction accuracy and improved spatial agreement, reducing classification errors.

Conclusions:

• The developed framework effectively reduces spatial disagreement among diverse ML-based FSMs.
• The optimized model offers improved reliability for flood susceptibility mapping.
• This approach can support the development of effective risk-based planning and early warning systems.