Content area
Floods are one of the most disastrous natural hazards ever in existence due to the immense effects that they have on land, human fatalities and buildings. Floods are a persistent challenge globally, which demand strategies that are effective for preparedness of disasters and mitigation of risks (Tehrani et al, 2015). The study focuses on the use of Machine learning (ML) models which are integrated with Geographic Information Systems (GIS) to identify high flood risk zones, on the basis of a detailed and comprehensive evaluation of susceptibility of flooding in Rathnapura, Sri Lanka. Rathnapura is vulnerable to recurrent flooding and has limited and scarce hydrological data. This study aims to overcome the limitations of traditional hydrodynamic approaches in such areas, with scarce hydrological data by use of Random Forest (RF), extreme Gradient Boosting (XGBoost), and Support Vector Machine (SVM), ML models to determine their efficiency in generating flood susceptibility maps.
Key factors affecting flooding conditions were selected for the study, including topography (elevation, aspect, slope), hydrological (rainfall, river proximity), an anthropogenic (land use, road proximity) variable. They were selected basing on availability of data and their influence. To assess the dependency of these factors on the occurrences of floods, a multicollinearity analysis was done using Variance Inflation Factor (VIF) and Mutual Information (MI) tests to streamline flood conditioning factors. Flood inventory dataset was obtained from past flood records and divided into training and testing datasets to optimize model accuracy through k-fold cross-validation. Flood susceptibility for each flood Machine learning model was classified into 5 risk levels from very low to very high, by use of natural breaks form of classification. Results indicated that RF was the best performing model since it had the highest accuracy in prediction on the internal and external datasets. Its robustness was displayed through outperforming the XGBoost and SVM in sensitivity, specificity, F1 – score and Cohen’s Kappa index. The study utilized Shapley Additive Explanations (SHAP) for feature importance analysis, which revealed that altitude, Topographic Roughness Index, rainfall, and proximity to rivers affect flood susceptibility significantly. The SHAP analysis revealed that factors such as Land use and soil type had a minimal influence at regional level but in localized flood events, their significance might be more pronounced.
The research highlights that ML can be used practically to generate flood risk maps as an alternative to traditional hydrological models in areas with limited data and also suggests the models to be transferred and utilized in other regions. As much as limitations exist, such as the inability of the models to quantify the flood depth and velocity, this research has provided insights into high – flood risk zones and this will enable informed resource allocation for the mitigation of floods. Future model research could explore model enhancements in additional factors an advanced selection of features to refine accuracy.