Abdo, Hazem Ghassan; Richi, Sahar Mohammed; Zerouali, Bilel; Alqadhi, Saeed; Katipoğlu, Okan Mert; Prasad, Pankaj; Arman, Hasan; Albanai, Jasem A; Mallick, Javed

Landslides significantly threaten natural and built environments, necessitating accurate prediction models for effective hazard mitigation. There is an urgent need to further improve the performance of machine learning algorithms in predicting landslide susceptibility by monitoring the impact of optimization algorithms on the performance of these models. This study evaluates the performance of various machine learning classifiers, including k-Nearest Neighbors (kNN), Multi-Layer Perceptron (MLP), and Extreme Gradient Boosting (XGBoost), for landslide susceptibility mapping. Additionally, Particle Swarm Optimization (PSO) is employed to enhance model performance by optimizing hyperparameters. Mountainous areas in the eastern Mediterranean (the northern Kabir River basin in western Syria) were identified as a result of the high frequency of landslide events over the past two decades. Nineteen factors causing landslides were identified, with no factor excluded, as a result of a multicollinearity test. The results indicate that XGBoost achieves the highest performance among traditional models. When integrated with PSO, the PSO-XGBoost model further improves classification performance, demonstrating its robustness in handling complex spatial patterns. Feature importance analysis using SHAP confirms slope as the dominant factor, followed by TRI, rainfall, Aspect, TWI, and curvature, highlighting the role of topography and hydrology in landslide occurrence. Moderate lithology, NDVI, and LULC contributions and lower importance of Flow Accumulation and Soil Depth suggest complex environmental interactions. Model predictions show varying susceptibility distributions. PSO-MLP assigns the highest very high susceptibility (44.09%), while PSO-XGBoost provides a balanced classification (31.13%). The PSO-XGBoost model demonstrates superior predictive capability, offering reliable landslide susceptibility maps for disaster risk management and land-use planning.

Achite, Mohammed; Katipoğlu, Okan Mert; Elshaboury, Nehal; Tuğrul, Türker; Pandey, Kusum

The accurate estimation of sediment discharge is crucial for the design and operation of engineering structures such as dams, water treatment facilities, and erosion control systems. This study evaluates the performance of various machine learning (ML) and deep learning (DL) models in predicting sediment transport in the Mina Basin, Algeria, at two stations: Oued Abtal and Sidi Abdelkader Djillali. The models include the sediment rating curve, category boosting, convolutional neural network, deep neural network (DNN), gated recurrent unit, and multilayer perceptron. Among these, the DNN model consistently demonstrated superior performance. For Oued Abtal station, the DNN achieved RMSE = 243.72 kg/s, MAE = 102.17 kg/s, NSE = 0.99, and PBIAS = 6.81%. At Sidi Abdelkader Djillali station, it recorded RMSE = 91.27 kg/s, MAE = 46.51 kg/s, NSE = 0.99, and PBIAS = 38.06%. Error analysis revealed that the DNN model offers the most reliable predictions, outperforming both traditional and other ML/DL methods. This study underscores the potential of deep learning models in advancing sediment transport prediction, particularly in semi-arid regions, and highlights their implications for sustainable water resource management.

Abdo, Hazem Ghassan; Richi, Sahar Mohammed; Prasad, Pankaj; Katipoğlu, Okan Mert; Halder, Bijay; Niknam, Arman; Hang, Hoang Thi; Alharbi, Maged Muteb; Mallick, Javed

Assessing landslide susceptibility is essential in urban planning and risk management. In the context of the Eastern Mediterranean region, there is a continuing need to compare the performance of machine learning (ML) algorithms in predicting landslide susceptibility, which can improve landslide risk management measures. Therefore, this study aims to evaluate and compare the predictive capabilities of three ML models: Multilayer perceptron (MLP), Light Gradient Boosting Machine (LGBM), and Extreme Gradient Boosting (XGBoost) models, in evaluating the susceptibility of various types of landslides and to refine the combination of causal factors. An evaluation of 19 conditioning factors, including topographical, geological, and environmental variables, was conducted to assess their effects on landslide susceptibility in different models in a geographic information system (GIS) environment. The results show that "Elevation" and "Slope" were consistently identified as the most influential factors in all models, with MLP demonstrating the greatest sensitivity to "Elevation." The study area was divided into five susceptibility categories: very low, low, moderate, high, and very high. According to the LGBM model, 24.27% of the area was classified as "very low" susceptibility, while the XGBoost and MLP models identified 25.69% and 27.28%, respectively. On the other hand, the "very high" susceptibility category covered 19.57%, 20.31%, and 19.78% of the area for the LGBM, XGBoost, and MLP models, respectively. The AUC-ROC approach has been utilized to evaluate, validate, and compare the performance of different ML models. Our study found AUC values for three MLTs. These findings suggest that all models demonstrate reasonable accuracy in identifying susceptible zones, and XGBoost demonstrated the best performance among the MLTs, with an AUC of 92.6% compared to the others. The insights gained from this study can inform targeted mitigation strategies to reduce landslide risks in Lebanon.

Katipoğlu, Okan Mert

The effective forecasting of droughts is the basic step of drought risk management, which consists of drought preparedness, mitigation, and early warning phases. This study used historical values of meteorological and hydrological drought indices to predict ERA 5 reanalysis data based on future hydrological droughts with a 1-month lead-time in Diyarbakır, in the Tigris basin. Regression tree (RT), boosted tree (BT), bagged tree (BAT), support vector machines (SVM), and adaptive neuro-fuzzy inference system (ANFIS) models and hybrid wavelet-machine learning models created by combining these models with discrete wavelet transform were analyzed to predict Standardized Runoff Index (SRI) based hydrological droughts. To construct the models and determine the temporal dependencies and relationships between variables, autocorrelation functions (ACF), partial autocorrelation functions (PACF), and correlation matrices were used. In addition, statistical criteria such as root mean square error (RMSE), mean absolute error (MAE) and, coefficients of determination (R 2), and the Taylor diagram were used to evaluate the success of the established models. The study has provided evidence that the application of wavelet transform can significantly improve the efficacy of hydrological drought prediction. The ANFIS model with Train RMSE: 0.65, MAE: 0.47, R 2: 0.49; Test RMSE: 0.84, MAE: 0.64, R 2: 0.39, which combines SRI (T − 1) and SRI (T) inputs, is considered the best stand-alone machine learning model for hydrological drought prediction. Furthermore, the most successful prediction performance is achieved by combining the SRI (T − 3), SRI (T − 2), SRI (T − 1), and SRI (T) inputs using the wavelet transform and BAT algorithm with Train RMSE: 0.59, MAE: 0.45, R 2: 0.59; Test RMSE: 0.74, MAE: 0.60, R 2: 0.48. Accordingly, it was concluded that hydrological droughts could be predicted using only previous hydrological droughts. Therefore, the study's outputs can be used to effectively plan water resources for decision-makers and various institutions in managing drought risk.