Modeling meteorological drought modeling for the Sakarya Basin

In this thesis study, a meteorological drought prediction modeling study was conducted for the Sakarya Basin using machine learning methods. A total of 13 stations were selected from the Sakarya Basin. The monthly average precipitation and temperature data of these stations for the years 1991-2023 were obtained from the Eskişehir Meteorology 3rd Regional Directorate. The Standardized Precipitation Evapotranspiration Index (SPEI) was used as the meteorological drought index. The SPEI values for 3-, 6-, 9-, and 12-month periods were calculated separately for each station. The drought prediction models of the meteorological stations were coded in Python using the Google Colab interface. The models employed included AdaBoost, Extreme Gradient Boosting (XGBoost), CatBoost, Multilayer Perceptron (MLP), Support Vector Regression (SVR), Linear Regression (LR), Extra Trees Method (ETM), Gaussian Process Regression (GPR), Gradient Boosting Method (GBM), Histogram-Based Gradient Boosting Regression (HGBR), Huber Method, K-Nearest Neighbors (KNN), LightGBM (LGBM), RANSAC, Random Forest (RF), and Ridge Regression machine learning methods. The performance of the models was evaluated using the coefficient of determination (R²) and mean squared error (MSE) metrics. It was found that the machine learning models provided better prediction results for the 12-month period compared to the other periods. In the prediction for the Polatlı station, the GPR model gave the best prediction results for the 12-month SPEI value with R²: 0.975 and MSE: 0.148. The temporal variations of the 12-month SPEI values were examined. The dates of the most severe droughts were identified as February 1995, October 2001, October 2007, November 2008, July 2013, December 2013, May 2021, April 2022, and March 2023. Drought maps were created separately for each station using the Inverse Distance Weighting (IDW) interpolation method based on the most severe drought dates. The drought prediction results obtained from the estimated station points on the IDW-generated maps were analyzed. The regions and stations experiencing severe droughts were identified through drought maps. To evaluate the most severe drought periods over a long time span, the 12-month SPEI values were compared with the values from IDW map points and machine learning models. Comparisons were made using MSE, root mean square error (RMSE), and mean absolute percentage error (MAPE) metrics. In machine learning models, the 12 month SPEI drought values were best approximated with MSE: 0.053, RMSE: 0.230, and MAPE: 5.230. In contrast, the estimated values from the IDW method yielded more distant predictions from the 12-month SPEI drought values, with MSE: 0.722, RMSE: 0.849, and MAPE: 33.324. A comprehensive drought modeling study was conducted by selecting at least one station from each sub-basin of the Sakarya Basin. In this study, the SPEI values for 3-, 6-, 9-, and 12-month periods were used as input data for each station. Different input combinations were created with the stations, and for the first time, SPEI was extensively evaluated in meteorological drought models using machine learning methods in the Sakarya Basin. Moreover, by employing machine learning methods such as HGBR, Huber, and Ridge, which had not been previously used for drought modeling in the Sakarya Basin, this study aims to provide a significant contribution to the literature.