Duvan et al. (2026) Enhancing Drought Prediction in Semi-Arid Climates: A Synthetic Data and Neural Network Approach Applied to Karaman Region, Turkey

Identification

• Journal: Atmosphere
• Year: 2026
• Authors: Akin Duvan, Sadık Alper Yıldızel
• DOI: 10.3390/atmos17020172

Research Groups

• Department of Civil Engineering, Engineering Faculty, Karamanoglu Mehmetbey University, Karaman, Turkey.

Short Summary

This study develops a drought forecasting framework for the semi-arid Karaman region of Turkey by combining synthetic data augmentation (KDE and Cholesky-based reconstruction) with Artificial Neural Networks (ANN). The approach successfully overcomes historical data scarcity, improving prediction accuracy for precipitation and drought intensity by 10–15% compared to traditional statistical models.

Objective

• To enhance long-term drought and precipitation prediction in data-scarce semi-arid climates by integrating synthetic data generation with interpretable neural network models.

Study Configuration

• Spatial Scale: Karaman Province, Turkey (Central Anatolia), a semi-arid region.
• Temporal Scale: Monthly resolution covering the period from January 1965 to December 2011 (46 years).

Methodology and Data

• Models used: Artificial Neural Networks (ANN) with a 14-neuron input layer and three hidden layers (64, 32, and 16 neurons); baseline comparisons using ARIMA (1,0,1) and Multiple Linear Regression (MLR).
• Data sources: Monthly precipitation records from the Turkish State Meteorological Service (MGM).
• Techniques:
  • Data Augmentation: Kernel Density Estimation (KDE) and Cholesky-based correlation reconstruction were used to triple the training dataset size.
  • Feature Engineering: Lagged variables (1, 2, and 3 months), rolling means (3 and 6 months), and cyclical temporal encoding (sine/cosine of months).
  • Interpretability: Local Interpretable Model-agnostic Explanations (LIME) to identify feature importance.
  • Validation: 80:20 train-test split, 5-fold cross-validation, and walk-forward validation.

Main Results

• Predictive Accuracy: The precipitation model achieved an $R^2$ of 0.76 and a Mean Absolute Error (MAE) of 12.8 mm. The drought intensity model achieved an $R^2$ of 0.72 and an MAE of 28.4%.
• Performance Gain: The synthetic-data-augmented ANN outperformed traditional methods and non-augmented ANNs by approximately 10–15%.
• Feature Influence: LIME analysis revealed that 1-month lagged precipitation and seasonal cyclical components were the most critical predictors for both models.
• Stability: Residual analysis (Durbin–Watson statistics near 2.0) and walk-forward validation confirmed the models' temporal stability and lack of systematic bias.

Contributions

• Methodological Innovation: Demonstrates a robust framework for using KDE and Cholesky decomposition to generate statistically consistent synthetic climate data for training deep learning models.
• Interpretability: Validates the use of LIME in hydroclimatic forecasting, providing transparency for "black-box" neural networks to support water management decisions.
• Regional Application: Offers a scalable and adaptable predictive tool specifically designed for the unique challenges of semi-arid Mediterranean ecosystems facing climate-induced water stress.

Funding

• This research received no external funding.

Citation

@article{Duvan2026Enhancing,
  author = {Duvan, Akin and Yıldızel, Sadık Alper},
  title = {Enhancing Drought Prediction in Semi-Arid Climates: A Synthetic Data and Neural Network Approach Applied to Karaman Region, Turkey},
  journal = {Atmosphere},
  year = {2026},
  doi = {10.3390/atmos17020172},
  url = {https://doi.org/10.3390/atmos17020172}
}

Generated by BiblioAssistant using gemini-3-flash-preview (Google API)