Aslan‐Sungur et al. (2025) Artificial neural networks estimate evapotranspiration for Miscanthus × giganteus as effectively as empirical model but with fewer inputs

Identification

• Journal: Theoretical and Applied Climatology
• Year: 2025
• Date: 2025-10-13
• Authors: Guler Aslan‐Sungur, Caitlin E. Moore, Carl J. Bernacchi, Emily A. Heaton, Andy VanLoocke
• DOI: 10.1007/s00704-025-05812-5

Research Groups

• Department of Agronomy, Iowa State University, Ames, Iowa, USA
• DOE Center for Advanced Bioenergy and Bioproducts Innovation, Iowa State University, Ames, Iowa, USA
• Department of Crop Sciences, University of Illinois, Urbana, Illinois, USA
• DOE Center for Advanced Bioenergy and Bioproducts Innovation, University of Illinois, Urbana, Illinois, USA
• School of Agriculture and Environment, The University of Western Australia, Crawley, Western Australia, Australia

Short Summary

This study compares artificial neural networks (FFN and NARX) with an empirical model (Granger and Gray) for estimating actual evapotranspiration (ET) of Miscanthus × giganteus. It demonstrates that ANNs can predict ET as effectively as empirical models but with fewer input variables, and the NARX model exhibits superior generalization across different sites.

Objective

• To test the performance of the Nonlinear Autoregressive Network with Exogenous Inputs (NARX) model in estimating actual evapotranspiration (ET) for Miscanthus × giganteus compared to other Artificial Neural Network (ANN) and empirical model approaches.
• To determine if ANNs can improve the accuracy, stability, and generalizability of ET estimates compared to empirical models while relying on fewer input variables.

Study Configuration

• Spatial Scale: Two agricultural research sites in the Central United States:
  • University of Illinois Energy Research Farm (UIEF) in Urbana, Illinois, USA (40°3′46″ N, 88°11′46″ W, approximately 220 m above sea level).
  • Sustainable Advanced Bioeconomy Research (SABR) Farm at Iowa State University, Ames, Iowa, USA (41°59′58″ N, 93°42′15″ W, approximately 314 m above sea level).
  • Focus on Miscanthus × giganteus crops.
• Temporal Scale:
  • UIEF: Nine-year eddy covariance dataset (2009–2017) used for training (2009–2016) and out-of-sample validation (2017).
  • SABR: Ten-month period (June 2019 to April 2020) used for out-of-sample testing.
  • Daily time step for ET prediction.

Methodology and Data

• Models used:
  • Artificial Neural Networks (ANNs):
    • Feed-Forward Neural Network (FFN)
    • Nonlinear Autoregressive Network with Exogenous Inputs (NARX)
  • Empirical Model: Granger and Gray (GG) model.
• Data sources:
  • Eddy covariance (EC) measurements for actual evapotranspiration (ETEC).
  • Supporting meteorological measurements: air temperature (Ta, in °C), relative humidity (Rh), incoming and outgoing short- and long-wave radiation, soil heat flux (G, in MJ m⁻² d⁻¹), soil temperature and moisture, photosynthetically active radiation (PAR), wind speed (u2, in m s⁻¹). Solar radiation (Rs, in W m⁻²) was estimated from PAR.
  • Data obtained from the Ameriflux network (US-UiB for UIEF, US-IAM for SABR).

Main Results

• The FFN model, using air temperature (Ta) and solar radiation (Rs) as inputs, demonstrated superior predictive power at the UIEF site (R² = 0.84, RMSE = 0.54 mm day⁻¹, MAE = 0.42 mm day⁻¹, Nash-Sutcliffe efficiency (NS) = 0.84 for out-of-sample validation).
• The NARX model performed optimally with only Ta as input (R² = 0.70, RMSE = 0.77 mm day⁻¹, MAE = 0.61 mm day⁻¹, NS = 0.68 for out-of-sample validation at UIEF).
• The empirical Granger and Gray (GG) model showed comparable performance to FFN at UIEF (R² = 0.83, RMSE = 0.94 mm day⁻¹, MAE = 0.76 mm day⁻¹, NS = 0.52).
• ANN approaches (FFN and NARX) achieved similar or better accuracy in estimating ET compared to empirical approaches, while requiring significantly fewer input variables (FFN: Ta, Rs; NARX: Ta; GG: Ta, Rh, Rs, u2, G).
• The NARX model exhibited superior generalization ability, maintaining its predictive accuracy when tested at a different site (SABR) (R² = 0.70, RMSE = 0.68 mm day⁻¹, MAE = 0.49 mm day⁻¹, NS = 0.71), outperforming FFN (R² = 0.70, RMSE = 0.69 mm day⁻¹, MAE = 0.58 mm day⁻¹, NS = 0.55) and GG (R² = 0.60, RMSE = 1.88 mm day⁻¹, MAE = 1.3 mm day⁻¹, NS = -1.24) at SABR.
• An optimal input data range of three years of historical data was determined to maximize the NARX model's prediction performance.
• Precipitation was found to have a minimal effect on ET predictions in this study.

Contributions

• This study is the first to comprehensively evaluate the NARX model for long-term (one year) actual evapotranspiration (ET) prediction at a daily time step for the bioenergy crop Miscanthus × giganteus.
• It provides a unique comparison of NARX and FFN ANN models against an established empirical model (Granger and Gray) for ET estimation in this specific cropping system.
• The research demonstrates that ANNs can achieve comparable or superior ET prediction accuracy with significantly fewer and more readily available input variables (air temperature and solar radiation, or just air temperature) compared to empirical models.
• It offers novel insights into the optimal input data range (three years) required for NARX model performance and assesses the stability and generalizability of ANN models across different geographical locations.

Funding

• DOE Center for Advanced Bioenergy and Bioproducts Innovation (U.S. Department of Energy, Office of Science, Biological and Environmental Research Program under Award Number DE-SC0018420).

Citation

@article{AslanSungur2025Artificial,
  author = {Aslan‐Sungur, Guler and Moore, Caitlin E. and Bernacchi, Carl J. and Heaton, Emily A. and VanLoocke, Andy},
  title = {Artificial neural networks estimate evapotranspiration for Miscanthus × giganteus as effectively as empirical model but with fewer inputs},
  journal = {Theoretical and Applied Climatology},
  year = {2025},
  doi = {10.1007/s00704-025-05812-5},
  url = {https://doi.org/10.1007/s00704-025-05812-5}
}