Katipoğlu, Okan Mert; Kartal, Veysi; Aktürk, Gaye; Kilinc, Huseyin Cagan; Elshaboury, Nehal; Yamany, Mohamed S.; Pande, Chaitanya Baliram

To achieve accurate and reliable predictions of groundwater drought, it is essential to incorporate relevant meteorological variables and apply advanced predictive methodologies capable of modeling their complex interrelationships. In this study, groundwater drought prediction models were developed based on the seven modified drought indices, which are obtained from groundwater levels and multiple meteorological parameters. The analysis focuses on four districts in Türkiye (Tekirdağ-Çorlu, Tekirdağ-Çerkezköy, Amasya-Merzifon, and Amasya-Centre), which are hereafter referred to collectively as the study areas. Three machine learning models were applied such as a standalone Artificial Neural Network (ANN) and two hybrid models, ANN integrated with Shuffled Frog Leaping Algorithm (SFLA-ANN) and Firefly Algorithm (FFA-ANN). The hybrid metaheuristic algorithms were chosen for their ability to model nonlinear and complex relationships between meteorological variables and groundwater levels, which traditional statistical methods cannot fully capture. Modified drought indices were employed to represent the groundwater drought better. The findings indicate that the hybrid FFA-ANN model generally provides the highest predictive accuracy for most indices, with coefficient of determination values frequently exceeding 0.90. In contrast, the ANN model occasionally outperforms the hybrid FFA-ANN model in terms of root mean square error (RMSE) for specific stations within the study areas. These results highlight the robustness of hybrid ANN approaches, especially FFA-ANN, in predicting groundwater drought under changing climate conditions. The novelty of this study is the modification of meteorological drought indices widely used to represent groundwater drought and the application of hybrid metaheuristic ANN models to estimate these indices in Türkiye. The results demonstrated that such models provide a reliable framework for predicting groundwater drought and provide decision-makers with advanced tools for water resources management, irrigation planning, and climate adaptation strategies.

Katipoğlu, Okan Mert; Terzioğlu, Zeynep Özge; Zerouali, Bilel

This study compares the performance of various models in predicting monthly maximum and average temperatures across three distinct regions: Samsun, Amasya, and Çorum. The evaluated models include Artificial Neural Network (ANN), Shuffled Frog Leaping Algorithm coupled with ANN (SFLA-ANN), Firefly Algorithm coupled with ANN (FFA-ANN), and Genetic Algorithm coupled with ANN (GA-ANN). In setting up the models, the dataset was divided into 70% for training and 30% for testing, and the outputs of the models were evaluated using various graphical and statistical indicators. The model with the smallest root mean square error (RMSE) value was selected for the maximum and average temperature predictions. Accordingly, for maximum and average temperature predictions, SFLA-ANN (RMSE of 2.93) and GA-ANN (RMSE of 3.55) in Samsun, GA-ANN (RMSE of 2.91) and GA-ANN (RMSE of 2.50) in Amasya and GA-ANN (RMSE of 2.97) and GA-ANN (RMSE of 2.50) in Çorum performed better than the other models, respectively. In addition, for the maximum temperature prediction with the highest accuracy, the R2 value of the SFLA-ANN model in Samsun was 0.89. In contrast, the R2 values of the GA-ANN model in Amasya and Çorum were determined as 0.91 and 0.91, respectively. Similarly, it was observed that the R2 values of the GA-ANN model for the average temperature prediction with the highest accuracy at Samsun, Amasya and Çorum stations were 0.78, 0.92 and 0.92, respectively. Overall, the GA-ANN consistently demonstrated superior performance in predicting both maximum and average temperatures across all three regions, as evidenced by its consistently low RMSE values. These findings provide valuable insights into selecting effective models for temperature prediction tasks in different geographical regions.

Achite, Mohammed; Emami, Somayeh; Emami, Hojjat; Katipoğlu, Okan Mert; Pandey, Kusum; Hajimirzajan, Amir; Elshaboury, Nehal

Effective water resource management in arid and semi-arid regions requires accurate estimation of hydrological parameters such as evaporation. This study investigates the monthly evaporation of the Sidi-M’Hamed Ben Aouda dam basin in northwest Algeria using six machine learning models: random forest (RF), extra tree (ET), gradient boosting (GB), category boosting (CatBoost), light gradient boosting machine (LGBM), and multi-layer perceptron (MLP). Climatic inputs (temperature, relative humidity, wind speed, and sunshine hours) from 1978 to 2023 were used to train and test the models. The findings reveal that the ET model achieved the best balance between accuracy and computational speed, while the RF model provided the highest overall accuracy. GB had a faster runtime with slightly reduced accuracy, whereas CatBoost and MLP underperformed. This comparative analysis highlights the suitability of ensemble tree-based models, particularly RF and ET, for accurate and efficient evaporation prediction, supporting water resource planning in data-scarce and climate-sensitive regions.

Koç, Erdal; Katipoğlu, Okan Mert

Within the scope of this study, a range of advanced machine learning and deep learning models—including Singular Spectrum Analysis (SSA), Adaptive Neuro-Fuzzy Inference System (ANFIS), Categorical Boosting (CatBoost), Convolutional Neural Network (CNN), Deep Autoencoder, Deep Neural Network (DNN), Gated Recurrent Unit (GRU), and Long Short-Term Memory (LSTM)—were employed to estimate the Standardized Groundwater Index (SGI) in Erzincan Province. SSA was utilized as a preprocessing technique to decompose input variables such as precipitation, relative humidity, temperature, and past SGI values into distinct components including trend, seasonality, cyclicality, and noise. These decomposed components were then fed into the artificial intelligence models to construct hybrid forecasting frameworks. The performance of each hybrid model was evaluated using multiple statistical indicators and visual analyses. The findings demonstrated that incorporating all SSA-derived subcomponents as inputs generally improved the monthly SGI prediction accuracy. However, for 12-month SGI predictions, the results were more variable, with both improvements and deteriorations observed depending on the model configuration. Additionally, the elimination of noise components was found to enhance both model generalization capability and overall prediction performance. Among the models tested, ANFIS emerged as the most effective in capturing GWD dynamics. To further investigate variable importance, Sobol sensitivity analysis was applied to the ANFIS outputs. The analysis revealed that previous SGI-1 values (t − 1) and relative humidity were the most influential inputs in predicting current SGI-1 (t) values.

Sarıgöl, Metin

The increase in average temperatures in the last century has caused global warming, which has increased the frequency of natural disasters. Floods are one of the most important natural disasters and harm the environment and especially human life. Flood routing techniques also play an important role in predicting floods. For this reason, the accuracy and precision of flood routing calculations are of vital importance in taking all necessary precautions before the floods reach the region and in preventing loss of life. This study aims to compare the performance of machine learning, deep learning and hybrid algorithms for flood routing prediction models in the Büyük Menderes River. In this research deep learning model Long-Short Term Memory (LSTM), machine learning model Artificial Neural Network (ANN), and hybrid machine learning models empirical mode decomposition (EMD)-ANN, and particle swarm optimization (PSO)-ANN algorithms were compared to forecast the flood routing results in two discharge observation stations in the Büyük Menderes river. The analysis results of the established ML algorithms were compared with statistical criteria such as mean error, mean absolute error, root mean square error and coefficient of determination. Additionally, Taylor diagrams, box plots, and beeswarm plot visual graphs were also used in this comparison analysis. At the end of the research, it was determined that the hybrid algorithm PSO-ANN was the most successful algorithm in forecasting flood routing results among other models according to the error values of MAE: 0.2514, MSE: 0.4613, RMSE: 0.6791, NSE: 0.941 and MBE: 0.047. Moreover, the LSTM algorithm was the approach with second estimation accuracy. The findings are vital in terms of taking necessary precautions and gaining time before floods reach any region.

Sarıgöl, Metin; Katipoğlu, Okan Mert; Dalkilic, Hüseyin Yildirim

Modeling monthly stream flows most accurately is of vital importance for water resource management, agricultural irrigation and efficient hydroelectric energy production, especially in semi-arid areas. Soft computing approaches have recently taken an important place in estimating streamflow time series. The potential of various data-driven approaches to predict streamflow in challenging climate conditions was evaluated. The study used machine and deep learning algorithms to model average monthly stream flows in two stream gauging stations in semi-arid region of the Konya closed basin where agriculture is at the forefront, accurate and reliable estimation of the stream flows is the basis of the study. For this, the performances of emotional neural network algorithm (EmNN), long-short term memory (LSTM), Elman neural network (ENN), nonlinear autoregressive exogenous model (NARX), recurrent neural network (RNN), group method of data handling (GMDH) were compared. The study’s unique contribution lies in its comprehensive comparison of these diverse algorithms, including newer approaches like EmNN, in the specific context of semi-arid hydrology. Partial autocorrelation analysis was applied to select input combinations, and lagged values exceeding 95% confidence limits were presented to the models as the most essential features. Artificial intelligence (AI) models use lagged stream flows to predict the streamflow time series. Statistical parameters, scatter diagrams and a time series approach are used to compare model performance. The GMDH model produced the following test results for 1604 no station: KGE: 0.656, R2: 0.608, NSE: 0.343, RMSE: 27.021, MAE: 3.834, MAPE: 0.662, MBE: −0.217, BF: 0.972. Similarly, for 1623 no station, the GMDH model yielded the following test results: KGE: 0.770, R2: 0.615, NSE: 0.531, RMSE: 0.006, MAE: 0.047, MAPE: 0.217, MBE: −0.012, BF: 0.956. In addition, the EmNN algorithm was the approach with second prediction accuracy. The findings of the study are important resources for optimizing the selection of AI models for streamflow prediction in semi-regional areas. The study also provides critical information for policymakers and decision-makers in similar climate zones worldwide for water resource management.