Predicting the Spatial Pattern of Flood Susceptibility Using Support Vector Machine Model in the Sirwan Watershed

Introduction and Goal
Floods cause significant financial losses and loss of life in the country every year. Although information on the location of flood events is of high scientific value, in many flood studies, flood zones have been determined solely based on expert’s opinion and multi-criteria decision-making methods. This study predicts the spatial pattern of flood susceptibility in the Sirwan watershed of Kurdistan province using spatial information on flood events in the last decade, and to achieve this objective, asupport vector machine model, as a machine learning-based approach, was employed.
Materials and Methods
In order to carry out this research, first a database in the geographic information system was prepared for flooding events in the Sirwan watershed using bank flood data from the Main Directorate of Natural Resources and Watershed Management of Kurdistan Province. The number of 102 flooding events in the Sirwan watershed was confirmed in the ten-year period (2015-2015). Since machine learning models require points of occurrence and non-occurrence of flooding, points of non-occurrence of flooding were also selected based on the homogeneous-unit method. This database was completed using information obtained from face-to-face interviews with local communities in this area. Based on the various characteristics of the Sirwan watershed and a review of scientific sources, sixteen factors affecting flooding events were selected and their digital maps were prepared. Factors affecting flooding included elevation, slope aspect, slope percentage, convergence index, drainage density, land use, maximum 24-hour precipitation, number of rainfalls higher than the average of the meteorological station, normalized vegetation difference index, plan curvature, profile curvature, soil texture, distance from the stream, topographic position index, topographic wetness index, and vertical distance from the stream, which were used as independent variables in the modeling. Flooding occurrence and non-occurrence data were randomly divided into two training and validation groups with proportions of 70% and 30%. After implementing the support vector machine model in the R software environment, a flood susceptibility map of the Sirwan watershed was prepared and the spatial pattern of flood susceptibility was examined. The accuracy of the aforementioned map was evaluated using the area under the curve (AUC) statistic of the receiver operating characteristic.
Results and Discussion
After validation, the results showed that the support vector machine model with an area under the receiver operating characteristic curve (AUC) of 0.921 (92.1%) has a high capability to predict flood-prone areas. Given that the model's prediction accuracy is reported to be more than 90%, based on the common classification of model efficiency, the performance of the support vector machine model in the Sirwan watershed is considered excellent. Based on the analyses, the very low, low, medium, high, and very high flood susceptibility classes include 51, 10, 17, 20, and 2% of the Sirwan watershed, respectively. Given the identification of flood-prone areas, the implementation priorities for the flood management plan have been clearly identified for implementing measures. Out of the 127 sub-watersheds of the Sirwan sub-watershed, 15 sub-watersheds are in the high flood-prone class and 8 sub-watersheds are in the very high flood-prone class. Population density is relatively high in all sub-watersheds in the high flood susceptibility category (such as Gazrokhani, Palangan, Shwishe, Sawji, Sleen, Sianaw, Danan, Zaribar, etc.) and in sub-watersheds in the very high flood susceptibility category (such as Sanandaj, Babarez, Marivan, Gholian, Mochesh, Doulbakh, Doroud, and Khamsan).
Conclusions and Suggestions
Based on the results of this study, the performance of the support vector machine model in identifying areas prone to flooding was very good. Based on the lack of specialized data and financial resources in the executive agencies, the application of this model in the management and planning of watershed improvement measures is important. In this study, prioritization of operational watersheds was performed based on the severity of flooding. Because, in conditions of data scarcity, it will not only save time and resources, but also improve the effectiveness of watershed management remedial measures. According to the results of this study, it is suggested that the support vector machine model be used at larger provincial, regional, and national levels, as well as in detailed-implementation studies of watershed management for risk management.