Extended Abstract
Introduction: Given the scarcity of data on river basins nationwide, numerous researchers turn to spatial analysis within a Geographic Information System (GIS) setting for hydrological studies and flood investigations. On that basis, identifying the most important factors influencing flood occurrence and severity, as well as building their sensitivity maps can be one of the most important solutions for flood reduction. Therefore, the objective of this study is to prepare a flood risk map in Bushehr province using machine learning techniques and to identify important factors affecting flood hazards.
Materials and methods: In this study conducted in Bushehr province, we aimed to compare the effectiveness of three machine learning models: Support Vector Machine (SVM), Random Forest (RF), and Generalized Additive Model (GAM). Initially, layers of information influencing flood occurrence in the study area were identified. Each prepared map served as input for the models. Various layers such as slope, slope direction, elevation, distance to river, drainage density, lithology, land use, topographic wetness index, and vegetation cover index were prepared using ArcGIS and SAGA-GIS software, crucial for analyzing flood patterns. Using data from 925 flood locations, points were divided into two sets: 70% (645 points) for modeling and 30% (280 points) for evaluation. The effectiveness of the models was validated using Receiver Operating Characteristic (ROC) analysis.
Results and Discussion: The results indicated that among the ten main factors, height, rainfall, and lithology were the most important factors affecting flood occurrence, while slope and distance from the river had the least impact. Evaluating model accuracy using ROC revealed very good accuracy for the SVM model (0.86), generalized additive model (0.85), and RF model (0.88). Flood sensitivity analysis showed RF and GAM methods identified the highest area in the low susceptibility class, while the SVM method identified the highest area in the medium susceptibility class. Results indicated that 37.32% of the study area had low sensitivity, 26.01% had medium sensitivity, 12.42% had high sensitivity, and 24.42% were very sensitive to flood hazards. Also, two other models have had very good accuracy for flood modeling in the studied area. The ROC related to the RF model, SVM, and generalized collective model showed an accuracy of 88.5 for the RF model, 86% accuracy for the SVM model, and 85% accuracy for the generalized collective model.
Conclusion: This study concludes that integrating machine learning models, namely SVM, RF, and GAM, with GIS analysis holds tremendous potential for advancing our understanding of flood patterns in Bushehr province. Leveraging these tools allows for a deeper comprehension of flood dynamics, aiding informed decision-making and effective mitigation strategies. This approach marks a significant leap forward in proactively addressing flood challenges and fostering resilient flood management practices in Bushehr province.