تهیه نقشه حساسیت سیل در حوزه شهری اقلید با استفاده از الگوریتم‌های یادگیری ماشینی

ABSTRACT
Flooding is among the most destructive natural hazards, causing substantial human casualties and economic losses worldwide each year. The preparation of flood risk maps and the identification of flood-prone areas play a critical role in reducing the potential damages associated with flooding. In this study, a flood sensitivity map was developed for the urban watershed of Eqlid in Fars Province using a range of flood conditioning factors, including slope, TRASP, topographic position index, stream power index, distance from the river, drainage density, surface curvature, rainfall, normalized difference vegetation index (NDVI), geology, land use, and soil type. Spatial data processing and factor preparation were conducted using ArcGIS Pro software. Subsequently, flood risk maps were generated using Random Forest (RF) and Support Vector Machine (SVM) machine learning algorithms implemented in the R programming environment, and the results of the two models were systematically compared. Model performance was evaluated using the receiver operating characteristic (ROC) curve, yielding accuracy values of 0.83 for the RF model and 0.84 for the SVM model, indicating good and nearly equivalent predictive performance. Feature importance analysis revealed that rainfall, soil type, land use, and drainage density were the most influential factors in the RF model, while rainfall, slope, soil type, and drainage density were identified as the dominant variables in the SVM model. The maps and datasets produced in this study can serve as valuable tools for urban planners and decision-makers in preventive planning, enhancing public awareness, and designing flood-resilient infrastructure.
Extended Abstract
Introduction
Flooding is one of the major natural hazards resulting from heavy rainfall, river overflows, and tidal influences. In recent years, increasing urbanization and climate change have intensified flood occurrences, particularly urban flooding, across many regions of the world. This phenomenon poses substantial social and economic risks, making accurate flood prediction essential for minimizing potential damages. Recent methodological advances, particularly in machine learning, have significantly enhanced the accuracy of flood hazard prediction. In this study, a flood hazard zoning map was developed for the urban watershed of Eqlid using advanced machine learning algorithms, namely Random Forest and Support Vector Machine. This approach integrates multi-source spatial data to support a comprehensive flood risk management strategy, with the potential to reduce damages associated with urban flooding.
 
Methodology
The study area covers approximately 40/623 hectares and has an elevation range from 1234 to 3213 meters above sea level, located in the southern part of Iran. Geographically, the watershed is located between 52°24'0" and 52°52'0" East longitude and between 30°43'40" and 31°4'20" North latitude. Initially, a flood point inventory map was prepared using 150 points, including both flood occurrence and non-occurrence locations, identified through field surveys, topographic maps, and data obtained from the General Directorate of Natural Resources of Fars Province. Additionally, 15 variables were selected for modeling based on methodologies commonly adopted in previous studies. These variables include elevation, slope, TRASP, drainage density, distance from the river, land use, geology, soil texture, Stream Power Index (SPI), Topographic Wetness Index (TWI), Topographic Roughness Index (TRI), Topographic Position Index (TPI), average annual rainfall, and the Normalized Difference Vegetation Index (NDVI). Multicollinearity among the selected variables was assessed using the Variance Inflation Factor (VIF) to identify potential linear correlations. Subsequently, machine learning algorithms, namely Random Forest (RF) and Support Vector Machine (SVM), were employed for flood hazard modeling. Finally, model performance was evaluated using statistical indices, including Accuracy, Sensitivity, Specificity, and the Kappa coefficient, in addition to the receiver operating characteristic (ROC) curve.
 
Results and discussion
Examining multicollinearity among the selected variables is essential for the reliable production of flood sensitivity maps. Among the 15 selected variables, elevation, TWI, and TRI were excluded because of high VIF values. The flood sensitivity maps indicate that areas with low to moderate susceptibility are predominantly located in the northern, northwestern, and southwestern parts of the watershed, whereas zones with high flood susceptibility are mainly concentrated in the central area.
Model performance evaluation showed that Sensitivity values for the training dataset were 0.93 for the RF model and 0.84 for the SVM model, while corresponding values for the validation dataset were 0.86 and 0.80, respectively. Specificity values reached 0.94 for RF and 0.83 for SVM in the training phase, and 0.86 and 0.73 in the validation phase. Accuracy values were 0.94 for RF and 0.89 for SVM for the training data, and 0.73 and 0.77 for the validation data, respectively. In addition, Kappa coefficients were 0.86 for RF and 0.73 for SVM in the training dataset, decreasing to 0.75 and 0.63 in the validation dataset. The AUC values for the RF model were 99% for the training data and 83% for the testing data, while the SVM model achieved AUC values of 93% and 84%, respectively, indicating high predictive accuracy for both models.
The variable importance analysis revealed that rainfall, soil type, land use, slope, and drainage density are the primary factors influencing flood occurrence, each contributing significantly to flood hazard formation. Although rainfall intensity tends to increase at higher elevations, the accumulation of runoff in lower elevation areas significantly elevates flood risk in these zones. Aridisols soils, characterized by low permeability and calcareous structure, promote increased surface runoff and accelerate flood occurrence. Land use types such as urban areas, agricultural lands, and moderately vegetated zones contribute to intensified flooding as a result of increased impervious surfaces and reduced infiltration capacity. Gentle slopes reduce flow velocity and prolong water retention time, thereby increasing the likelihood of flooding. Although high drainage density can facilitate rapid water conveyance, it also concentrates runoff in downstream areas, potentially leading to sudden and destructive flood events.
These findings are consistent with previous studies and highlight the importance of considering the combined effects of natural and anthropogenic factors in urban planning and water resource management to mitigate flood hazards. Overall, the machine learning models employed in this study produced accurate and reliable flood hazard maps that can effectively support urban management decisions and contribute to reducing flood-related damages in the urban watershed of Eqlid.
 
Conclusion
In this study, advanced machine learning models, including Random Forest and Support Vector Machine, were employed to generate flood hazard potential maps for the urban watershed of Eqlid with a high level of accuracy. The findings indicate that the interaction of climatic, geological, and hydrological variables, together with land use changes and urbanization expansion, plays a fundamental role in flood occurrence. The proposed models demonstrated a strong capability to accurately predict flood sensitivity and to provide valuable information for decision-makers involved in flood damage reduction planning. However, certain limitations, including the focus on a single watershed and constraints related to data availability, should be taken into account in future research. Overall, this research emphasizes that the application of data-driven methods and modern modeling approaches, alongside scientific analyses, can serve as an effective strategy for reducing the destructive impacts of floods in urban areas and supporting policymakers in informed decision-making.
 
Funding
There is no funding support.
 
Authors’ Contribution
Conceptualization and formal analysis and initial methodology and investigation by S.Negahban.; project administration and supervision and visualization and software writing review and editing by M.Marhamat.;
 
 
Conflict of Interest
Authors declared no conflict of interest.
 
Acknowledgments
We are grateful to all the scientific consultants of this paper.