Sadeghzadeh et al. (2026) Deep learning fusion modeling of reference evapotranspiration with multi-source remote sensing data through addressing noise impacts

Identification

• Journal: Smart Agricultural Technology
• Year: 2026
• Date: 2026-02-01
• Authors: Mostafa Sadeghzadeh, Sepideh Karimi, Sungwon Kim, Jalal Shiri, Il-Moon Chung
• DOI: 10.1016/j.atech.2026.101891

Research Groups

• Department of Water Engineering, Faculty of Agriculture, University of Tabriz, Tabriz, Iran
• Water Engineering and Science Research Institute (WESRI), University of Tabriz, Tabriz, Iran
• Department of Railroad Construction and Safety Engineering, Dongyang University, Yeongju, 36040, Republic of Korea
• Department of Land, Water and Environmental Research, Korea Institute of Civil Engineering and Building Technology, Goyang-si, 10223, Republic of Korea

Short Summary

This study developed a deep learning-based Convolutional Neural Network (CNN) fusion method to estimate daily reference evapotranspiration (ETo) using multi-source remote sensing data, specifically evaluating its performance under noisy input conditions. The Random Forest model coupled with CNN fusion (RF-CNN) significantly outperformed other fusion and direct methods in accuracy and stability across both humid and arid regions, even with added Gaussian noise.

Objective

• To develop a robust and scalable deep learning-based fusion method for integrating multi-source satellite-based high-resolution data (MODIS, ERA5, CHIRPS) to estimate daily reference evapotranspiration (ETo).
• To systematically evaluate the performance of this fusion method and other pixel-level fusion techniques under both clean and noisy input data conditions.
• To assess the impact of individual input variables on ETo simulation using these fusion methods through interpretability analysis.

Study Configuration

• Spatial Scale: Two distinct regions in Iran (Mazandaran province for humid climate and Bushehr province for arid climate). All datasets were resampled to a uniform 1 km grid.
• Temporal Scale: 23 years (2000–2023) of daily satellite and ground-based data.

Methodology and Data

• Models used:
  • Deep Learning Fusion: Convolutional Neural Network (CNN), Autoencoder (AE).
  • Machine Learning Regression: Random Forest (RF).
  • Traditional Fusion: Weighted Average (WA), Principal Component Analysis (PCA), Kalman Filter (KF).
  • Benchmark/Empirical ETo: FAO Penman-Monteith (FPM-56), Hargreaves-Samani (HGR), Modified Hargreaves (MHGR), Makkink, Mass transfer method.
• Data sources:
  • Satellite: MODIS (Normalized Difference Vegetation Index (NDVI), albedo), ERA5 (solar radiation), CHIRPS (precipitation).
  • Observation: Ground-based meteorological stations (daily minimum and maximum air temperatures (Tmin, Tmax), relative humidity (RH), wind speed (Ws)).
  • Noise Simulation: Gaussian noise was added to input data at 10%, 20%, and 30% of each input's standard deviation.

Main Results

• The RF model coupled with the CNN-fusion method (RF-CNN) was the most accurate and stable approach for regional ETo estimation under both clean and noisy data conditions.
  • For clean data, RF-CNN achieved R² values of 0.990 and 0.997, and RMSE values of 0.177 mm/day and 0.092 mm/day for the humid (north) and arid (south) regions, respectively.
• The proposed CNN-fusion method significantly improved RF model performance, reducing the Scatter Index (SI) by an average of 56% compared to the direct method under clean conditions.
• While Gaussian noise escalated error magnitudes for all models, RF-CNN demonstrated superior robustness, maintaining the highest accuracy even with noisy inputs.
• SHAP analysis revealed distinct influential factors for ETo in different climates:
  • In the humid northern region, maximum air temperature (Tmax) and solar radiation (Rs) were the most influential parameters.
  • In the arid southern region, solar radiation (Rs) and wind speed (Ws) were the most crucial variables.
• Calibration of empirical equations generally improved their performance and increased their robustness to noisy data.
• Cross-regional validation showed a substantial decrease in model performance (R² of 0.803 for North-to-South and 0.851 for South-to-North), indicating regional specificity of the model.

Contributions

• Development of a novel deep learning-based 1D Convolutional Neural Network (CNN) for pixel-level fusion of multi-source satellite data (MODIS, ERA5, CHIRPS) to estimate daily ETo, capturing hierarchical and nonlinear patterns.
• Systematic comparison of the proposed CNN-fusion method with other pixel-level fusion techniques (Weighted Average, Principal Component Analysis, Kalman Filter, Autoencoder) under both clean and noisy data conditions.
• Integration of fusion outputs with Random Forest (RF) regression and SHAP analysis to provide accurate and interpretable ETo estimations, identifying feature importance.
• Explicit quantification and assessment of noise impacts on fusion-based ETo models, addressing a previously unassessed aspect in the literature.

Funding

• Institutional Research Program of the Korea Institute of Civil Engineering and Building Technology (KICT) under the project “Demonstration Technology Development for Integrated Operation of Subsurface and Sand-Filled Dams (20250442-001)”.

Citation

@article{Sadeghzadeh2026Deep,
  author = {Sadeghzadeh, Mostafa and Karimi, Sepideh and Kim, Sungwon and Shiri, Jalal and Chung, Il-Moon},
  title = {Deep learning fusion modeling of reference evapotranspiration with multi-source remote sensing data through addressing noise impacts},
  journal = {Smart Agricultural Technology},
  year = {2026},
  doi = {10.1016/j.atech.2026.101891},
  url = {https://doi.org/10.1016/j.atech.2026.101891}
}