Multi-Class Flood Classification Model Using Spatial Topo-Hydrological Features and Interpretable Machine Learning

Abstract

Rapid urban expansion of cities in many developing countries, such as Nigeria, is aggravating occurrences of devastating urban floods because of sudden changes in climate and uncontrolled land use. Methods based on traditional flood prediction are expensive, primarily binary-classification-based, and lacking generality, which hampers their capability for disaster preparedness purposes. In this paper, we propose a standardized multiclass flood classification framework with spatial, topographic, hydrological and meteorological covariates based on three ML classifiers, including Random Forest, Support Vector Machine and Logistic Regression. The model was evaluated strictly using stratified 5-fold cross-validation and a 20% held-out test set. From the right, the RF model recorded the highest performance accuracy, at 92%, indicating desirable generalisation and resistance to overfitting. SVM was successful with 87% and LR achieved 83%, both being relatively unstable at minor flood classification. The experimentation-based feature importance analysis indicated that the environmental data index is the most important predictor, increasing interpretation and transparency in modelling. The results introduce RF as a trustworthy multi-class urban flood classification tool in data-poor contexts, with potential applications for early warning systems, city management and climate-resilient policies in the Global South.