Next generation data-driven flood susceptibility modelling with spatial machine learning

Abstract

Accurate flood susceptibility assessment remains a critical challenge, yet spatial machine learning offers next-generation data-driven solutions for robust and scalable flood risk prediction. Traditional flood susceptibility models based on hydrodynamic and statistical approaches are often constrained by extensive data requirements, complex calibration, and high computational costs, which limit their application in data-scarce regions. This study advances existing approaches by integrating Multi-Criteria Decision Analysis (MCDA) and the Analytic Hierarchy Process (AHP) with spatial machine learning to optimize the weighting of flood conditioning factors prior to model training. Expert and literature derived weights for nine spatial predictors were normalized and used as input layers to three algorithms—Random Forests (RF), Support Vector Machines (SVM), and Convolutional Neural Networks (CNN)—for flood susceptibility mapping in Chinhoyi, Zimbabwe. A total of 564 flood and 925 non-flood locations were mapped using CNN, 432 flood and 564 non-flood locations using RF, and 569 flood and 908 non-flood locations using SVM. Model performance was assessed using accuracy metrics and receiver operating characteristics to determine predictive capability and generalization. Results revealed that CNN outperformed RF and SVM, producing superior spatial precision and reliability. The methodological integration of AHP-MCDA with deep spatial learning represents a novel advancement in flood susceptibility modelling, enhancing model generalization, interpretability, and applicability in data-limited environments. The study contributes to the advancement of geospatial artificial intelligence applications in hydrological hazard modelling, offering practical insights for resilient urban planning, early warning systems, and sustainable disaster risk management in vulnerable landscapes.