Ndulue et al. (2025) Machine learning-based estimation of daily reference evapotranspiration across agro-ecological zones in Nigeria: comparative analysis and model ranking

Identification

• Journal: Theoretical and Applied Climatology
• Year: 2025
• Date: 2025-11-26
• Authors: Emeka Ndulue, Ikenna Onyekwelu, Kelechi Igwe, Vintus Ogwo, Okechukwu Michael
• DOI: 10.1007/s00704-025-05902-4

Research Groups

• Department of Agricultural & Bioresources Engineering, University of Nigeria, Nsukka, Nigeria
• Agriculture and Natural Resource Management, Schreiner University, Kerrville, TX, USA
• Carl and Melinda Helwig Department of Biological and Agricultural Engineering, Kansas State University, Manhattan, KS, USA

Short Summary

This study evaluated five machine learning models for daily reference evapotranspiration (ETo) estimation across six agro-ecological zones in Nigeria under various data availability scenarios, demonstrating their potential for reliable ETo estimation with minimal inputs, particularly the Bagging model, to enhance water resource management.

Objective

• To evaluate and rank the performance of five machine learning models for daily reference evapotranspiration (ETo) estimation across six agro-ecological zones in Nigeria under 15 different input data combinations.
• To identify machine learning models that can reliably estimate ETo with fewer meteorological inputs, comparable to the FAO Penman-Monteith model, for improved water resource and irrigation management in Nigeria.

Study Configuration

• Spatial Scale: Six agro-ecological zones in Nigeria (Port Harcourt, Ibadan, Enugu, Kaduna, Sokoto, Maiduguri), covering a land area of approximately 923,769 km². Climate data were sourced at a spatial resolution of 0.5° × 0.5°.
• Temporal Scale: Daily reference evapotranspiration (ETo) estimation over a 34-year period (1990–2023).

Methodology and Data

• Models used: Bootstrap Aggregating (Bagging), Linear Regression (LR), Multilayer Perceptron (MLP), Random Forest (RF), and M5P Pruning (M5P). The FAO Penman-Monteith (FAO PM) equation served as the standard for comparison.
• Data sources: National Aeronautics and Space Administration Prediction of Worldwide Energy Resource, v2.4.15 (NASA POWER) satellite imagery, derived from NASA’s GEOS 5.12.4 FP-IT and GMAO MERRA-2. Meteorological parameters included solar radiation (Rs), maximum air temperature (Tmax), minimum air temperature (Tmin), relative humidity (RH), and wind speed at 2 m height (U2).

Main Results

• Under complete data conditions, all machine learning models demonstrated high accuracy (Nash-Sutcliffe Efficiency (NSE): 0.924–0.999; Root Mean Square Error (RMSE): 0.038–0.446 mm/day during testing).
• Machine learning model performance varied significantly by location and input data combinations, showing distinct trends between northern and southern Nigerian regions.
• The overall ranking of the machine learning models, based on a multi-metric approach, was: Bagging > M5P > MLP > RF > LR.
• In data-limited scenarios:
  • For southern Nigeria, models using solar radiation (Rs) and relative humidity (RH) as inputs were most effective, with ETo estimates being less accurate without Rs data.
  • For northern Nigeria, models using relative humidity (RH) and wind speed (U2) as inputs yielded better results, indicating their critical influence on ETo in these regions.
• The Bagging model consistently exhibited the best performance across most scenarios and locations, attributed to its ensemble learning approach.

Contributions

• Provided the first comprehensive evaluation and ranking of five machine learning models for daily ETo estimation across six diverse agro-ecological zones in Nigeria under 15 different data input combinations.
• Identified optimal minimal input data combinations for accurate ETo estimation in distinct northern (relative humidity and wind speed) and southern (solar radiation and relative humidity) Nigerian regions.
• Offered practical guidance for irrigation planning, design, and water management in Nigeria, particularly in regions with limited meteorological data availability.
• Utilized NASA POWER data to address the challenge of sparse distribution and incomplete climatic records from ground weather stations in Nigeria.

Funding

• NASA Langley Research Center POWER Project, funded through the NASA Earth Science/Applied Science Program.

Citation

@article{Ndulue2025Machine,
  author = {Ndulue, Emeka and Onyekwelu, Ikenna and Igwe, Kelechi and Ogwo, Vintus and Michael, Okechukwu},
  title = {Machine learning-based estimation of daily reference evapotranspiration across agro-ecological zones in Nigeria: comparative analysis and model ranking},
  journal = {Theoretical and Applied Climatology},
  year = {2025},
  doi = {10.1007/s00704-025-05902-4},
  url = {https://doi.org/10.1007/s00704-025-05902-4}
}