Novel forecasting approaches using combination of machine learning and statistical models for flood susceptibility...

Area of Science:

• Environmental science
• Geospatial analysis
• Machine learning applications

Background:

• Flood susceptibility assessment is crucial for disaster management.
• Individual machine learning models show variable predictive performance.
• Ensemble methods can improve prediction stability and accuracy.

Purpose of the Study:

• To implement and compare eight individual machine learning models for flood susceptibility.
• To introduce and evaluate seven novel ensemble models for enhanced flood prediction.
• To identify the most accurate and reliable model for flood susceptibility assessment.

Main Methods:

• Implementation of artificial neural networks, classification and regression trees, flexible discriminant analysis, generalized linear model, generalized additive model, boosted regression trees, multivariate adaptive regression splines, and maximum entropy.
• Development of seven ensemble models: EMca, EMciInf, EMciSup, EMcv, EMmean, EMmedian, and EMwmean.
• Validation using data from 201 flood events in the Haraz watershed, Iran, and 10,000 non-occurrence points.

Main Results:

• Boosted regression trees achieved the highest Area Under the Receiver Operating Characteristic (AUROC) among individual models (0.975).
• The proposed Ensemble Model median (EMmedian) yielded the highest accuracy (0.976) across all tested models.
• Significant variability was observed in predictions from individual models, highlighting the need for ensemble approaches.

Conclusions:

• Ensemble forecasting approaches, particularly the EMmedian, are recommended for flood susceptibility assessment.
• The EMmedian model offers improved generalizability, stability, and reduced sensitivity compared to individual models.
• Utilizing ensemble methods effectively reduces uncertainty in flood susceptibility predictions.