Investigating the Role of the Key Conditioning Factors in Flood Susceptibility Mapping Through Machine Learning Approaches

This study harnessed the formidable predictive capabilities of three state-of-the-art machine learning models-extreme gradient boosting (XGB), random forest (RF), and CatBoost (CB)-applying them to meticulously curated datasets of topographical, geological, and environmental parameters; the goal was to investigate the intricacies of flood susceptibility within the arid riverbeds of Wilayat As-Suwayq, which is situated in the Sultanate of Oman. The results underscored the exceptional discrimination prowess of XGB and CB, boasting impressive area under curve (AUC) scores of 0.98 and 0.91, respectively, during the testing phase. RF, a stalwart contender, performed commendably with an AUC of 0.90. Notably, the investigation revealed that certain key variables, including curvature, elevation, slope, stream power index (SPI), topographic wetness index (TWI), topographic roughness index (TRI), and normalised difference vegetation index (NDVI), were critical in achieving an accurate delineation of flood-prone locales. In contrast, ancillary factors, such as annual precipitation, drainage density, proximity to transportation networks, soil composition, and geological attributes, though non-negligible, exerted a relatively lesser influence on flood susceptibility. This empirical validation was further corroborated by the robust consensus of the XGB, RF and CB models. By amalgamating advanced deep learning techniques with the precision of geographical information systems (GIS) and rich troves of remote-sensing data, the study can be seen as a pioneering endeavour in the realm of flood analysis and cartographic representation within semiarid fluvial landscapes. The findings advance our comprehension of flood vulnerability dynamics and provide indispensable insights for the development of proactive mitigation strategies in regions that are susceptible to hydrological perils.