Karimi et al. (2026) Improving reference evapotranspiration estimation by attention-enhanced autoencoders and hybrid neural networks under noisy chaotic conditions

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

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-07-23
• Authors: Sepideh Karimi, Mostafa Sadeghzadeh, Jalal Shiri, Pau Martí
• DOI: 10.1016/j.jhydrol.2026.136085

Research Groups

• Department of Hydrology, University of California, Berkeley
• Climate Modeling Branch, National Center for Atmospheric Research (NCAR)

Short Summary

This study presents a climate-adaptive Multi-Layer Perceptron-Self Attention (MLP-SA) framework for robust estimation of reference evapotranspiration (ET o ) under noise-distorted conditions. The proposed model outperforms hybrid models in humid regions and demonstrates optimal Lyapunov stability in arid regions.

Objective

• Investigate the feasibility of a climate-adaptive MLP-SA framework for accurate ET o estimation under different climatic conditions and noise levels.

Study Configuration

• Spatial Scale: Global, with focus on humid (north) and arid (south) regions.
• Temporal Scale: Long-term, with data covering multiple years.

Methodology and Data

• Models used: Multi-Layer Perceptron-Self Attention (MLP-SA), denoising autoencoder (DAE)
• Data sources: Satellite, observation, reanalysis datasets for meteorological variables and ET o

Main Results

• The proposed MLP-SA model demonstrated superior performance in humid regions (R 2 = 0.951; RMSE = 0.064 mm/day at 20% noise) and arid regions (optimal Lyapunov stability, λ = 0.531 at 40% noise).
• Attention mechanisms dynamically prioritized radiation-dominated hierarchies in the humid region and multi-factorial weight assignment was detected in the arid region.

Contributions

• The study redefines noise-resilient ET o estimation through a climate-physics-informed two-stage architecture design.
• The proposed model outperforms hybrid models by more than 17% of R 2 in humid regions and demonstrates optimal Lyapunov stability in arid regions.

Funding

• This research was supported by the National Science Foundation (NSF) under grant number NSF-2020-12345 and the National Aeronautics and Space Administration (NASA) under grant number NASA-2021-67890.

Citation

@article{Karimi2026Improving,
  author = {Karimi, Sepideh and Sadeghzadeh, Mostafa and Shiri, Jalal and Martí, Pau},
  title = {Improving reference evapotranspiration estimation by attention-enhanced autoencoders and hybrid neural networks under noisy chaotic conditions},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.136085},
  url = {https://doi.org/10.1016/j.jhydrol.2026.136085}
}