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

Extended Abstract

Introduction

Flood is one of the significant natural hazards caused by heavy rainfall, river overflows, and tidal waves. In recent years, with increasing urbanization and climate change, especially urban floods, its intensity has escalated in various regions around the world. This phenomenon brings significant social and economic risks, and accurate prediction of it is essential to reduce damages. Modern methods, including machine learning, have been able to increase the accuracy of flood hazard prediction. In this study, an effort has been made to prepare a flood hazard zoning map in the urban watershed of Eqlid by utilizing advanced machine learning algorithms (Random Forest and Support Vector Machine). This approach integrates multi-source data to provide a comprehensive flood risk management strategy, which can be effective in reducing damages caused by urban floods.



Methodology

The study area covers approximately 40/623 hectares with an elevation range between 1234 to 3213 meters above sea level, located in the southern part of Iran.

This watershed is geographically located between 52°24'0" to 52°52'0" East longitude and 30°43'40" to 31°4'20" North latitude. Initially, a flood point map was prepared based on 150 points, consisting of both flood occurrence and non-occurrence points, which were identified through field visits, topographic maps, and data obtained from the General Directorate of Natural Resources of Fars Province. Additionally, 15 variables were selected for modeling, based on the approach used by previous researchers. These variables include elevation, slope, trasp, drainage density, distance from 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 NDVI. A multicollinearity analysis was performed using the Variance Inflation Factor (VIF) to determine linear correlations between the variables. Subsequently, machine learning algorithms, including Random Forest (RF) and Support Vector Machine (SVM), were used for modeling. Finally, the accuracy of the models was evaluated using statistical indices such as Accuracy, Sensitivity, Specificity, and Kappa, as well as the ROC curve.



Results and discussion

The examination of multicollinearity among the selected variables for producing flood sensitivity maps is of great importance. Among the 15 chosen variables, Elevation, Topographic Wetness Index (TWI), and Topographic Roughness Index (TRI) were excluded due to high Variance Inflation Factor (VIF) values. The sensitivity maps indicate that areas with low to moderate sensitivity are mainly located in the northern, northwestern, and southwestern parts, whereas highly flood-prone zones are primarily situated in the central part of the watershed. In the performance evaluation of the models, Sensitivity values for the training data were 0.93 for RF and 0.84 for SVM, and for the validation data, 0.86 for RF and 0.80 for SVM. Specificity values were 0.94 for RF and 0.83 for SVM in training data, and 0.86 for RF and 0.73 for SVM in validation data. The Accuracy values obtained were 0.94 for RF and 0.89 for SVM in training data, and 0.73 for RF and 0.77 for SVM in validation data. Finally, Kappa values were 0.86 for RF and 0.73 for SVM in training data, and 0.75 for RF and 0.63 for SVM in validation data. The AUC values in the RF model were 99% for training and 83% for testing data, and in the SVM model, 93% and 84%, respectively, indicating the high accuracy of both models. The results of the variable importance analysis showed that rainfall, soil type, land use, slope, and drainage density are the main influencing factors in flood occurrence, each playing a key role in the formation of flood hazards. Rainfall is higher at higher elevations, but runoff accumulation in lower elevation areas increases the flood risk in these regions. Aridisols soil, characterized by low permeability and calcareous structure, leads to increased surface runoff and accelerates flood occurrence. Land uses such as urban areas, agriculture, and moderate vegetation contribute to intensifying flooding due to increased impervious surfaces and reduced water infiltration. Low slope decreases water flow velocity and prolongs water retention time, increasing the likelihood of flooding. Although high drainage density facilitates rapid water discharge, it concentrates runoff in downstream areas, potentially causing sudden and destructive floods. These findings are consistent with previous studies and emphasize the necessity of considering the combination of natural and human factors in urban planning and water resource management to reduce flood hazards. Overall, the machine learning models used in this study were able to produce accurate and reliable flood hazard maps that can be practical for urban management decisions and reducing flood damages in the urban watershed of Eqlid.



Conclusion

In this study, advanced machine learning models such as Random Forest and Support Vector Machine were utilized to produce flood hazard potential maps in the urban watershed of Eqlid with high accuracy. The findings indicate that the combination of climatic, geological, and hydrological variables along with land use changes and urbanization growth plays a fundamental role in flood occurrence. The presented models demonstrated the capability to accurately predict flood sensitivity and provide valuable information for decision-makers in managing damage reduction programs. However, limitations such as the study's focus on a specific watershed and data scarcity exist, which should be considered in future research development. Ultimately, this research emphasizes that the use of data-driven methods and modern models alongside scientific analyses can be an effective strategy to reduce the destructive consequences of floods in urban areas and assist policymakers in optimal decision-making.