Are Presence-Only Machine Learning Models Reliable Enough for Hazard Mapping? Insights from Flood Hazard in the Maghreb

Abstract

Classification-based machine learning (ML) models have gained significant popularity for environmental hazard mapping, using the spatial distribution of both presence and absence locations to classify areas at risk. However, a critical challenge arises in accurately identifying “absence points” in hazard mapping. Absence points may not truly reflect regions where a natural disaster cannot occur; rather, they may represent areas with insufficient historical records or limited monitoring, leading to potential biases in model outputs. To address this issue, presence-only ML models are particularly useful as they focus on locations where floods have been observed, eliminating the need for accurately identified absence data. In this study, three presence-only models, i.e., One-Class Support Vector Machine (OneclassSVM), Maximum Entropy (MaxEnt), and Genetic Algorithm for Rule-set Production (GARP), were evaluated to map flood hazard in the Maghreb region. The results show that MaxEnt outperformed the other models with an accuracy of 96.8 % and an area under the curve (AUC) of 0.978, followed by GARP with 94.928% accuracy and 0.96 AUC, and SVM with 92.42% accuracy and 0.93 AUC. These results bring attention to the potential of MaxEnt model in flood hazard assessment and prediction in cases where only presence data are available.