Sadeghzadeh et al. (2026) Interpretable Temperature‐Based Deep Learning for Evapotranspiration: SHAP ‐Based Feature Analysis in CNN ‐ GPU

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Meteorological Applications
• Year: 2026
• Date: 2026-01-01
• Authors: Mostafa Sadeghzadeh, Jalal Shiri, Sepideh Karimi, Ozgur Kisi
• DOI: 10.1002/met.70148

Research Groups

Not explicitly stated in the abstract.

Short Summary

This study developed and evaluated deep learning models (CNN-RNN and CNN-GPU) for estimating reference evapotranspiration (ETo) using only temperature data, demonstrating that the CNN-GPU model achieved high accuracy and computational efficiency while implicitly capturing the influence of other meteorological variables.

Objective

• To estimate temperature-dependent reference evapotranspiration (ETo) using deep learning models (CNN-RNN and CNN-GPU) when comprehensive meteorological data inputs are limited.

Study Configuration

• Spatial Scale: Not explicitly stated, but the mention of 'study region' and 'spatial meteorological patterns' implies a regional or local scale.
• Temporal Scale: Not explicitly stated, but the Root Mean Square Error (RMSE) reported in 'mm/day' suggests a daily or sub-daily temporal resolution for the estimations.

Methodology and Data

• Models used: CNN-RNN, GPU-accelerated CNN (CNN-GPU). The FAO-Penman-Monteith (FPM-56) equation is mentioned as a classical approach for context. SHapley Additive exPlanations (SHAP) analysis was used for model interpretability.
• Data sources: Limited input variables: temperature records. Other meteorological data (solar radiation, wind speed) were implicitly captured by the models, as revealed by SHAP analysis.

Main Results

• The CNN-GPU model outperformed the CNN-RNN model in both accuracy and computational efficiency.
• CNN-GPU model accuracy: Root Mean Square Error (RMSE) = 0.23 mm/day, Nash-Sutcliffe efficiency (NS) = 0.98.
• CNN-GPU model computational efficiency: 20.4% faster training time compared to CNN-RNN.
• SHAP analysis revealed that solar radiation and wind speed exerted high degrees of implicit influence on the models, even when excluded from the direct input matrix, clarifying complex nonlinear relationships.
• The proposed deep learning models successfully captured complex temporal and spatial meteorological patterns in the study region despite being trained with limited input variables (temperature records).

Contributions

• Demonstrates the feasibility and high performance of deep learning models (specifically CNN-GPU) for accurate reference evapotranspiration (ETo) estimation using only temperature data, addressing data scarcity issues.
• Provides insights into the implicit capture of complex nonlinear relationships between ETo and other meteorological variables (solar radiation, wind speed) by deep learning models, even when these variables are not direct inputs.
• Introduces a computationally efficient (20.4% faster training) and accurate CNN-GPU model as a viable alternative for ETo estimation.
• Utilizes SHAP analysis to enhance the interpretability of deep learning models in hydrological applications, revealing hidden dependencies.

Funding

Not explicitly stated in the abstract.

Citation

@article{Sadeghzadeh2026Interpretable,
  author = {Sadeghzadeh, Mostafa and Shiri, Jalal and Karimi, Sepideh and Kisi, Ozgur},
  title = {Interpretable Temperature‐Based Deep Learning for Evapotranspiration: <scp>SHAP</scp> ‐Based Feature Analysis in <scp>CNN</scp> ‐ <scp>GPU</scp>},
  journal = {Meteorological Applications},
  year = {2026},
  doi = {10.1002/met.70148},
  url = {https://doi.org/10.1002/met.70148}
}