نوع مقاله : پژوهشی
نویسندگان
1 دانشیار گروه پژوهشی مهندسی حفاظت خاک و آب، پژوهشکده حفاظت خاک و آبخیزداری، سازمان تحقیقات، آموزش و ترویج کشاورزی، تهران، ایران
2 دانشگاه آزاد اسلامی
چکیده
کلیدواژهها
موضوعات
عنوان مقاله [English]
نویسندگان [English]
Introduction and Goal
Flooding is one of the most destructive natural disasters with social, economic and environmental consequences, and machine learning methods have been developed to model and predict it. The purpose of this research is to zone and predict flood risk using three deep learning models, including GRU, LSTM and TCN models, and to introduce the most suitable ones in the Qarasu basin of Golestan province. The most important innovation of the present research, compared to previous research on hazard zonation in the country, is to use new deep learning models and consider as many influence factors on flood occurrence as possible in order to determine the most efficient model and increase the accuracy of flood prediction maps.
Materials and Methods
First, the research area was selected, and then the maps of the effective factors were collected and prepared. In this study, 16 effective factors were selected as independent variables. Using information received from the Ministry of Energy and field survey as the dependent variable, a flood event distribution map was prepared and divided into two groups of experimental points (30%) and training points (70%), randomly. TCN, LSTM and GRU models were implemented, flood hazard zonation maps were prepared and classified into five classes: very low, low, medium, high and very high. The classification accuracy and validation of the flood risk zoning and prediction maps were evaluated. Finally, the most appropriate model was selected.
Results and Discussion
In order to prepare a flood event zonation map, 59 flood locations were prepared based on available information and radar images before and after the flood using Google Earth software (dependent variable). Of these, 70% were randomly selected and divided as training data and 30% as test data for model implementation and validation, respectively. In the next step, 16 geological, hydrological, and morphometric group factors of the basin were used along with climatic data (as independent variables) for zoning and modeling. Evaluation of the classification accuracy of the models using two indices, frequency ratio (FR) and seed cell area index (SCAI), showed that most flood locations are in high-hazard classes (high and very high) and these zones occupy a larger share of the area. Of course, this value was lower in the GRU model. Also, using the area under curve of ROC (Receiver Operating Characteristic) and CC (Cost Curve), the TCN model is given priority over the two LSTM and GRU models. It has the highest (0.92) and lowest (0.08) values among the models, respectively. It was also found that in the TCN model, the very high hazard class with an area of about 45% of the region covers about 83% of the flood events, in the LSTM model with an area of about 23% of the region, it covers about 60% of the flood events, and in the GRU model with an area of about 10% of the region, it covers about 4% of the flood events. Therefore, the flood hazard zonation maps obtained from this study can be the basis for planning and crisis management caused by flood events. The results of this study will be essential for future development projects of various organizations active in many developing countries and will help as a basis to reduce flood risk and manage it.
Conclusion and Suggestions
After running the models using the area under the ROC curve (AUC), all three models had high modeling suitability in the training phase with an excellent score (0.9-1). The maps prepared from flood hazard zonation using all three models indicated high accuracy in zone classification and appropriate distribution of flood points in high and very high classes. However, the TCN model achieved good modeling fit in both training and testing phases (above 0.9). Evaluation of classification accuracy and validation of the models showed that the two TCN and LSTM models, in addition to having appropriate class thresholds in classification, have higher priority for zonation and flood prediction. Finally, using the Area Under the Cost Curve, values of 0.08, 0.10, and 0.11 were obtained for the TCN, LSTM, and GRU models, respectively, which indicated that the TCN model with the lowest cost has high modeling utility. Given that morphometric indices play an important role in the zoning of flood hazard maps, the use of these parameters along with other conventional indices in the preparation of zonation maps is recommended in future studies. Also, given that deep learning methods and their combination have yielded good results in studies abroad, it is recommended that interested researchers investigate and evaluate new hybrid methods and the use of appropriate optimizers in other sub-basins of Golestan province.
کلیدواژهها [English]
)