Abubakar et al. (2025) Evaluation of temporal convolutional networks and ensemble machine learning models for meteorological drought prediction in the Nigerian Sudano–Sahelian zone

Identification

• Journal: Discover Environment
• Year: 2025
• Date: 2025-11-19
• Authors: Muhammad Lawal Abubakar, Auwal F. Abdussalam, Zaharaddeen Isa, Muhammad Sambo Ahmed, A. Musa, Jonah Birga, Mohammed Ismail
• DOI: 10.1007/s44274-025-00456-8

Research Groups

• Department of Geography and Sustainability Studies, Kaduna State University, Kaduna, Nigeria
• Climate Research Group, Kaduna State University, Kaduna, Nigeria
• Department of Geography, Federal University Dutsin-Ma, Katsina, Nigeria

Short Summary

This study predicted and characterized meteorological drought in the Sudano–Sahelian region of Nigeria (SSRN) using temporal convolutional networks (TCNs) and ensemble machine learning models. It revealed severe droughts in the 1970s and 1980s, identified dominant drought cycles linked to large-scale climatic oscillations, and demonstrated the high predictive power and generalization consistency of TCNs for drought forecasting.

Objective

• To predict and characterize meteorological drought in the Sudano–Sahelian region of Nigeria (SSRN).
• To evaluate the performance of temporal convolutional networks (TCNs) and ensemble machine learning models (linear regression, AdaBoost, XGBoost) for meteorological drought prediction.
• To extract cyclical patterns of droughts using fast Fourier transform (FFT).
• To detect regions facing water stress and enhance water management and mitigation techniques by analysing the geographical and temporal distributions of drought characteristics.

Study Configuration

• Spatial Scale: Sudano–Sahelian Region of Nigeria (SSRN), located between latitudes 10° and 14°N and longitudes 4°E and 14°E. Data from 11 synoptic meteorological stations across the region.
• Temporal Scale: 1972–2022 (51 years) for rainfall and temperature time series. Drought analysis and prediction were performed using the Standardized Precipitation Evapotranspiration Index (SPEI) at a 12-month timescale (SPEI-12).

Methodology and Data

• Models used:
  • Standardized Precipitation Evapotranspiration Index (SPEI-12) for drought determination.
  • Hargreaves method for Potential Evapotpiration (PET) calculation.
  • Temporal Convolutional Networks (TCNs) for drought prediction.
  • Linear Regression (LR) as a baseline model for drought prediction.
  • AdaBoost Regressor as an ensemble baseline model for drought prediction.
  • Extreme Gradient Boosting (XGBoost) as an ensemble baseline model for drought prediction.
  • Fast Fourier Transform (FFT) for extracting cyclical patterns of droughts.
  • Run theory for characterizing drought events (duration, severity, intensity).
  • Min-max scaler for data normalization.
  • Adam optimizer for TCN training with Mean Squared Error (MSE) loss and Mean Absolute Error (MAE) metrics.
  • GridSearchCV with fivefold time series cross-validation for hyperparameter tuning of AdaBoost and XGBoost.
• Data sources:
  • Daily rainfall, minimum temperature (Tmin), and maximum temperature (Tmax) from 11 synoptic meteorological stations, obtained from the Nigerian Meteorological Agency (NiMet).
  • NINO3.4 index from the National Oceanic and Atmospheric Administration (NOAA) Climate Prediction Center.

Main Results

• Severe meteorological droughts were identified in the 1970s and 1980s across the SSRN.
• Drought characteristics varied spatially:
  • Number of drought events ranged from 16 to 28, with Sokoto having the highest (28) and Borno/Kano the lowest (16).
  • Average drought duration ranged from 5.4 months to 9.2 months, with Borno having the longest (9.2 months) and Sokoto the shortest (5.4 months).
  • Average drought severity ranged from -7.9 to -13.8 (SPEI units), with Borno exhibiting the highest severity (-13.8).
  • Average drought intensity ranged from -1.24 to -1.34 (SPEI units per month), with Borno having the highest intensity (-1.34).
• Fast Fourier Transform (FFT) analysis revealed dominant drought cycles of 24 months (corresponding to quasi-biennial oscillation), 72 months (El Niño–Southern Oscillation, ENSO), and 144 months (Atlantic Multidecadal Oscillation, AMO).
• Model performance for SPEI-12 prediction:
  • Linear Regression (LR) achieved the highest mean coefficient of determination (R²) of 98% and the lowest mean absolute error (MAE) across most stations.
  • Temporal Convolutional Networks (TCNs) achieved a mean R² of 97% and a mean Critical Severity Index (CSI) of 0.765.
  • TCNs reduced the MAE by 36–48% compared with AdaBoost and by 28–40% relative to XGBoost across most stations.
  • While LR occasionally recorded marginally lower MAEs and slightly higher R² values, TCNs exhibited greater generalization consistency across spatial domains, indicating better adaptability to complex climatic variability.
  • Ensemble models (AdaBoost and XGBoost) performed poorly in high-drought zones.

Contributions

• This study offers a novel hybrid methodology by integrating Temporal Convolutional Networks (TCNs), Fast Fourier Transform (FFT), and run theory for forecasting and characterizing the Standardized Precipitation–Evapotranspiration Index (SPEI) in the Sudano–Sahelian Region of Nigeria (SSRN).
• It leverages TCN's superior ability to model long-term temporal dependencies within sequential climate data and FFT's effectiveness in recognizing periodic signals and climatic oscillations associated with drought-inducing events.
• The research enhances both the accuracy and interpretability of drought prediction models, providing valuable insights for stakeholders in food production, water resource governance, ecosystem management, and disaster risk management.
• It highlights the importance of regional context in model selection, demonstrating that both linear and deep learning models can achieve high predictive accuracy for drought forecasting in semiarid environments, depending on signal complexity and predictor interactions.

Funding

This research did not receive any funding.

Citation

@article{Abubakar2025Evaluation,
  author = {Abubakar, Muhammad Lawal and Abdussalam, Auwal F. and Isa, Zaharaddeen and Ahmed, Muhammad Sambo and Musa, A. and Birga, Jonah and Ismail, Mohammed},
  title = {Evaluation of temporal convolutional networks and ensemble machine learning models for meteorological drought prediction in the Nigerian Sudano–Sahelian zone},
  journal = {Discover Environment},
  year = {2025},
  doi = {10.1007/s44274-025-00456-8},
  url = {https://doi.org/10.1007/s44274-025-00456-8}
}