Physics-informed neural networks for enhanced reference evapotranspiration estimation in Morocco: Balancing...

Area of Science:

• Environmental Science
• Agricultural Engineering
• Artificial Intelligence

Background:

• Reference evapotranspiration (ETo) is crucial for agriculture, but standard methods like the Penman-Monteith equation are data-intensive.
• Existing simplified models have limitations, and purely data-driven AI methods raise concerns about explainability and reliability.
• There is a need for accurate, trustworthy, and data-efficient ETo estimation techniques.

Purpose of the Study:

• To evaluate the integration of semi-physical (SP) models into the loss function of Physics-Informed Neural Networks (PINNs) for ETo estimation.
• To assess the impact of combining data-driven and physics-based losses on model performance across various data availability scenarios.
• To enhance the explainability and reliability of AI models for ETo estimation in environmental applications.

Main Methods:

• Developed a novel PINN approach combining a data-driven loss with losses from SP models (Priestley-Taylor, Makkink, Hargreaves-Samani, Abtew) using a convex combination parameter (θ).
• Collected in-situ agrometeorological data (air temperature, solar radiation, relative humidity, wind speed) from Morocco.
• Calibrated SP models using CMA-ES optimization and trained/evaluated the PINN under different data scenarios and θ values.

Main Results:

• The integrated PINN approach significantly improved ETo estimation accuracy (RMSE, R²) compared to fully data-driven or baseline models across all tested data scenarios.
• The PINN required less training time to reach optimal values when θ was within the [0.2, 0.8] range.
• Optimal θ values were determined for each SP model and data scenario, demonstrating the flexibility of the approach.

Conclusions:

• Physics-Informed Neural Networks offer a promising, accurate, and trustworthy method for reference evapotranspiration estimation.
• Integrating physical constraints into AI models enhances their reliability and applicability in operational environmental management.
• This study represents a significant step towards developing controlled, physics-informed AI for environmental science applications.