نوع مقاله : مقاله پژوهشی
موضوعات
عنوان مقاله English
نویسندگان English
Aims: Floods are among the primary natural hazards that cause economic and social damage. The Ghare Ghom watershed in Razavi Khorasan province is prone to dangerous floods due to its climatic conditions, topography, and physiographic features. Most rivers in the basin are seasonal, leading to occasional severe flooding. This study assessed flood susceptibility mapping in this region.
Materials & Methods: Machine learning models, including the Random Forest (RF) and Classification and Regression Tree (CART), were used for flood susceptibility mapping. Of the 117 recorded flood events, 70% were used for training and 30% for validation. Influential factors included land use, slope degree, geology, distance from river, digital elevation model, slope direction, geomorphology, soil type, plan and profile curvature, rainfall, and the Topographic Wetness Index (TWI). Model performance was evaluated using the Area Under the ROC Curve (AUC) and the Tolerance Index.
Findings: The RF model results show that distance from the river, rainfall, and altitude had the greatest impact on flood sensitivity. In the CART model, altitude, distance from the river, and rainfall were the most influential factors. The CART model classified the area into very low, low, medium, high, and very high susceptibility classes as 9%, 29.8%, 21.9%, 32.4%, and 6.9%, respectively. For the RF model, these values were 11.48%, 24.8%, 28.7%, 24%, and 11%.
Conclusion: The Area Under the Curve (AUC) was 0.91 for the CART model and 0.87 for the RF model. The prediction rate was 0.88 for CART and 0.83 for RF. These results indicate that the CART model performed better than the RF model in predicting flood susceptibility.
Innovation: Flood susceptibility mapping and identifying influential factors using machine learning are key contributions of this study. The results support planners and policymakers in managing flood risks and reducing future economic and social losses in the region.
کلیدواژهها English
Floods are one of the main natural hazards that cause economic and social damage. Also, they cause the most damage and deaths in the world. Floods occur due to the high water level and the overflow of rivers and their flow into plains and residential areas. In addition, natural disasters destroy natural resources, roads, the economy, agricultural lands and people's lives. Floods have increased with the expansion of urban areas, the sharp increase in population and the destruction of forest lands. The Ghare Ghom watershed in Razavi Khorasan province is one of the basins where, due to climatic conditions, topography, and physiographic characteristics, as well as the rivers that mostly have seasonal water, sometimes dangerous floods occur in the region. Therefore, this study assessed flood susceptibility mapping in the Ghare Ghom watershed in Razavi Khorasan province.
Identifying and collecting information about the factors affecting flooding is the first stage of zoning studies, and selecting important factors plays a major role in the accuracy of these maps. Therefore, in this research, machine learning models, including random forest model (RF) and classification and regression tree model (CART), were used for flood susceptibility mapping. Of the 117 flood events, 70% were considered for training and 30% for validation. Land use, degree of slope, geology, distance from Revier, digital elevation measure, direction, geomorphology, soil science, land curvature shape, land curvature profile, rainfall and topographic wetness index (Topographic Wetness Index) were considered as effective factors. Finally, the area under the ROC curve and the Tolerance index were used to evaluate the models.
The results of multicollinearity between independent variables were examined using the well-known statistical indices of collinearity (VIF) and Tolerance (Tol). Multicollinearity analysis was performed on the 12 risk predictors used in this study for flood susceptibility. The results indicate that the VIF and ToL of all factors are less than the threshold. Therefore, all factors were entered into the modeling process for further analysis. RF model indicate that the variables of distance from the river, rainfall, and altitude, and in the CART model, altitude, distance from the river, and rainfall had the most significant impact on flood sensitivity. The areas covered by very low, low, medium, high, and very high classes in the CART model are 9, 29.8, 21.9, 32.4, and 6.9 percent, respectively. These results for the RF model are 11.48, 24.8, 28.7, 24 and 11 percent, respectively. The area under the curve (AUC) in the CART model is equal to 0.91, and the RF model is equal to 0.87. So, based on the prediction rate, the CART model was equal to 0.88, and the RF model was equal to 0.83. The results show that the CART model performs better than the RF model. In both sensitivity maps, the very high sensitivity class is located in the western and southwestern parts of the basin, full of urban and rural areas. As a result, flooding causes extensive damage to urban and rural areas, areas prone to the passage of power transmission lines and routes, road construction, industries, pastures, agricultural lands, etc. In both models, the flood sensitivity map showed that the sensitivity to flooding is higher in downstream areas than upstream and high-altitude areas due to lower elevation and slope. Also, in downstream areas, due to severe land use changes in recent years, the sensitivity to damage has increased, resulting in a decrease in vegetation cover and an increase in agricultural and residential lands. The area under the curve (AUC) in the CART model is equal to 0.91, and the RF model is equal to 0.87. Thus, based on the prediction rate, the CART model was equal to 0.88, and the RF model was equal to 0.83. The results show that the CART model performs better than the RF model.
The modelling results showed that the CART model has a higher efficiency than the RF model. Overall, the results indicate that more than half of the study area is in the high and very high-risk sensitivity class. There are 30 urban points and more than 1000 rural points in the area, most of which are located in high-risk areas and are exposed to the highest risk. Most of these residential areas are located next to and within the boundaries of the basin's waterways. This study found that the analysis of environmental factors, using remote sensing data and spatial modelling, is a powerful tool for predicting sensitive areas. Given the high accuracy of machine learning models, especially the CART model in this study, and given the lack of hydrological data in most of the country's watersheds and the excellent application of spatial modeling for natural disaster management in the country, it is suggested that researchers use this model to prepare flood risk potential maps in other regions of the country and compare the results with other models This research helps planners and policymakers in the management of natural disasters to identify flood susceptibility areas and reduce the future economic and financial losses caused by them.
The manuscript did not receive a grant from any organization
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
