Flood susceptibility mapping: Integrating machine learning and GIS for enhanced risk assessment

Abstract

Flooding presents a formidable challenge in the United States, endangering lives and causing substantial economic damage, averaging around $5 billion annually. Addressing this issue and improving community resilience is imperative. This project employed machine learning techniques and publicly available data to explore the factors influencing flooding and to develop flood susceptibility maps at various spatial resolutions. Six machine learning algorithms, including Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), K-nearest neighbor (KNN), Adaptive Boosting (Ada Boost), and Extreme Gradient Boosting (XGB) were used. Geospatial datasets comprising thirteen predictor variables and 1528 flood inventory data collected since 1996 were analyzed. The predictor variables are rainfall, elevation, slope, aspect, flow direction, flow accumulation, Topographic Wetness Index (TWI), distance from the nearest stream, evapotranspiration, land cover, impervious surface, land surface temperature, and hydrologic soil group. Five hundred twenty-eight non-flood data points were randomly created using a stream buffer for two scenarios. A total of 2964 data points were classified into flooded (1) and non-flooded (0) categories and used as a target. Overall, testing results showed that the XGB and RF models performed relatively well in both cases over multiple resolutions compared to other models, with an accuracy ranging from 0.82 to 0.97. Variable importance analysis depicted that predictor variables such as distance from the streams, hydrologic soil type, rainfall, elevation, and impervious surfaces significantly affected flood prediction, suggesting a strong association with the underlying driving process. The improved performance and the variation of the susceptible areas across two scenarios showed that considering predictor variables with multiple resolutions and appropriate non-flooding training points is critical for developing flood-susceptibility models. Furthermore, using tree-based ensemble algorithms like RF and XG boost in the stack generalization approach can help achieve robustness in a flood susceptibility model where multiple algorithms are being evaluated.