Geospatial Artificial Intelligence (GeoAI) Framework for Flood Susceptibility Mapping Using a Hybrid SVM–XGBoost Model

Abstract

The Gharb Plain, a vital socioeconomic and agricultural region in northwestern Morocco, faces persistent threats from catastrophic flooding driven by its low-lying topography and complex hydrological networks. Conventional flood assessment methodologies often fail to capture the highly non-linear environmental dynamics of the basin or suffer from a lack of model interpretability. This study addresses these gaps by developing an advanced, high-resolution GeoAI framework designed to accurately map flood susceptibility and support sustainable land-use planning in the region. A novel two-stage hybrid model was developed by coupling eXtreme Gradient Boosting (XGBoost) for probabilistic feature extraction with a Support Vector Machine (SVM) acting as a meta-classifier. To rigorously evaluate its performance, the SVM–XGB model was benchmarked against three advanced architectures: a Multilayer Perceptron (MLP), a One-Dimensional Convolutional Neural Network (CNN1D), and ShallowNet. All models were trained on 80% of the dataset and validated on the remaining 20%. Post-hoc model interpretability and feature importance were quantified using SHapley Additive exPlanations (SHAP) to align algorithmic decisions with physical watershed mechanics. Quantitative evaluation demonstrated the clear superiority of the hybrid SVM–XGBoost model, which achieved a testing accuracy of 0.9895, a Cohen’s Kappa of 0.9789, and a testing AUC of 0.9809, effectively eliminating the overfitting tendencies and spatial fragmentation observed in the CNN1D network. SHAP analysis revealed that the elevation is the primary driver of flood susceptibility, followed closely by soil permeability and drainage density. Spatially, the SVM–XGBoost map delivered the most cohesive boundaries, precisely delineating highly vulnerable areas within the central merjas zone (inundated depression) and the critical Oued Sebou and Oued Beht confluence near Kenitra. The findings confirm that integrating tree-based boosting with structural risk minimization allows the hybrid framework to effectively map complex, non-linear flood pathways with high generalization capability. By bridging the gap between “black box” machine learning and actionable hydrological science via SHAP, this framework provides a highly trustworthy decision-support tool. The resulting high-resolution susceptibility maps offer a robust, empirical foundation for local authorities to enforce strict zoning regulations and strategically implement nature-based solutions designed to simultaneously mitigate surface flood peaks.