Geographic information systems-based hydrological basin modeling and hydrograph prediction using machine learning methods

Water is an indispensable resource for living organisms. Initially, people settled near water sources to meet their needs. However, with the growing population and the expansion of agricultural activities, the demand for water increased, leading to the development of various transmission and storage methods. Today, factors such as population growth, industrialization, increased consumption, and global warming are causing the depletion of clean water resources. Consequently, the comprehensive and sustainable management of river basins has become essential. Predicting river hydrographs is crucial for enabling effective planning and management of river basins. In this study, hydrological watershed modeling of the Ağva, Ulupınar, Çandır, Doyran, and Karaman streams located in Antalya was carried out using the CBS-based program NetHydro. First, ASTER GDEM data was downloaded, a digital elevation model was created, and watershed boundaries were determined based on AGI locations. The Thiessen polygon method was used to determine precipitation areas, and rainfall data from observation stations were tested using the Kolmogorov-Smirnov test, and the precipitation amounts falling on the watershed were calculated. Flood discharges for the rivers were calculated using the DSİ synthetic, SCS, Snyder, and Mockus methods. Additionally, flood discharges were statistically predicted using AGI data, and results were compared with previously obtained flood discharges from earlier studies. In the Western Mediterranean and Antalya Watersheds, it was understood that statistical methods could not be used effectively, as the results were far from actual values. It was observed that the DSİ synthetic and SCS unit hydrograph methods reflected the regional characteristics, and the results were consistent with previous studies. The NetHydro program is considered to provide consistent results, facilitate hydrological studies, and offer a practical alternative. Flow forecasting was carried out using the LSTM architecture, a deep learning method. Flow forecasts were made for the Ağva, Ulupınar, Çandır, Doyran, and Karaman streams using AGI data. In the LSTM architecture, optimization methods such as Adam, RMSprop, and SGDM were used, and the memory length of the model was determined to be 30, 60, 90, and 120 days. Winsorization was applied at 0%, 2.5%, and 5% levels to reduce the effect of outliers. NSE, R², KGE, MSE, RMSE, and MAE performance metrics were used to assess the model's performance. Upon examination of the results, it was found that the best results were obtained with Adam, followed by RMSprop and SGDM. In terms of daily memory, results for 60 and 90 days were generally better compared to others, but there was no complete consistency. Increasing the Winsorization ratio improved the model performance and reduced the impact of outliers. The Winsorization process resulted in an R² value ranging from 80% to 97%. This shows that the model performance is directly related to the data structure and that the model's performance decreases when predicting outliers. The LSTM model was very successful in flow forecasting, and it was concluded that its weaknesses should be addressed through hybrid models or new approaches. In conclusion, this thesis emphasizes the reliability of the DSİ and SCS methods in flood discharge calculations, while also demonstrating that the LSTM model is an effective method for flow prediction. The study provides both theoretical and practical contributions to hydrological modeling and flood management.