Sabegh et al. (2025) Enhancing reference evapotranspiration prediction with biological ensemble support vector regression and MODIS data integration

Identification

• Journal: Sustainable Water Resources Management
• Year: 2025
• Date: 2025-12-15
• Authors: Sanaz Monavvar Sabegh, Davoud Zarehaghi, Saeed Samadianfard
• DOI: 10.1007/s40899-025-01317-1

Research Groups

• Department of Soil Science and Engineering, Faculty of Agriculture, University of Tabriz, Tabriz, Iran
• Department of Water Science and Engineering, Faculty of Agriculture, University of Tabriz, Tabriz, Iran
• Water Sciences and Hydroinformatics Research Center, Khazar University, Baku, Azerbaijan

Short Summary

This study developed a novel Biological Ensemble Support Vector Regression (BE-SVR) model, integrating meteorological and MODIS remote sensing data, to enhance reference evapotranspiration (ET0) prediction in semi-arid regions, demonstrating superior accuracy compared to conventional and optimized SVR models.

Objective

• To develop and evaluate a Biological Ensemble Support Vector Regression (BE-SVR) model for daily reference evapotranspiration (ET0) estimation in the semi-arid Tabriz region of Iran.
• To assess the contribution of remotely-sensed indices and various input scenarios to ET0 prediction.
• To evaluate the BE-SVR model's performance against a benchmark Support Vector Regression (SVR) and a Firefly Algorithm-optimized SVR (FFA-SVR), using FAO-56 Penman–Monteith ET0 estimates for validation.

Study Configuration

• Spatial Scale: A specific semi-arid agricultural region in Tabriz, East Azerbaijan Province, northwestern Iran (38.12° N, 46.24° E, elevation: 1365 m).
• Temporal Scale: Daily data from 2003 to 2023 (20 years), split into training (2003–2016) and testing (2017–2023) periods.

Methodology and Data

• Models used:
  • Biological Ensemble Support Vector Regression (BE-SVR)
  • Firefly Algorithm-optimized Support Vector Regression (FFA-SVR)
  • Support Vector Regression (SVR) (baseline)
  • FAO-56 Penman–Monteith (FPM) equation (for reference/validation)
  • Kalman filtering (for Land Surface Temperature (LST) gap filling)
  • Cubic spline interpolation (for converting multi-day MODIS indices to daily resolution)
• Data sources:
  • Meteorological data: Daily climate data (2003–2023) from the Iran Meteorological Organization, including minimum, maximum, and mean air temperature (Tmin, Tmax, Tm), minimum, maximum, and mean relative humidity (RHmin, RHmax, RHm), wind speed at 2 m (U2), and sunshine duration (n).
  • Remote sensing data: MODIS Aqua satellite data (2003–2023) accessed via Google Earth Engine, including Land Surface Temperature (LST - daytime, nighttime, mean), Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Leaf Area Index (LAI), and Fraction of Photosynthetically Active Radiation (FPAR).

Main Results

• The BE-SVR model achieved exceptional accuracy in daily ET0 prediction, with its best configuration (Scenario 15) yielding a Root Mean Square Error (RMSE) of 0.159 mm day⁻¹, Mean Absolute Error (MAE) of 0.108 mm day⁻¹, Coefficient of Determination (R²) of 0.998, Nash–Sutcliffe Efficiency (NSE) of 0.998, and Willmott’s Index of Agreement (WI) of 0.999 during the test period.
• BE-SVR consistently outperformed FFA-SVR (best RMSE = 0.227 mm day⁻¹, R² = 0.995) and baseline SVR (best RMSE = 0.467 mm day⁻¹, R² = 0.980). This represents a 66.6% reduction in RMSE compared to baseline SVR and a 29.9% improvement over FFA-SVR.
• Wind speed (U2) and relative humidity (RH) were identified as dominant meteorological drivers, significantly improving model accuracy.
• The integration of MODIS-derived remote sensing variables (LST and vegetation indices) into BE-SVR (Scenario 15 vs. 16) substantially improved performance, reducing RMSE by 52.7% compared to using meteorological inputs alone.
• At monthly and seasonal temporal scales, FFA-SVR showed slightly better performance in capturing short-term dynamics. However, BE-SVR demonstrated greater capacity in capturing interannual variability and maintaining generalization across multiple years.

Contributions

• Development of a novel, biologically inspired ensemble machine learning framework, Biological Ensemble Support Vector Regression (BE-SVR), for ET0 estimation.
• Introduction of an aggregation-free ensemble architecture that combines kernel-space representations of multiple SVRs trained on distinct feature subsets, improving generalization and avoiding output-weighting bias.
• Demonstrated the synergistic fusion of ground-based meteorological data and MODIS-derived remote sensing indices for enhanced ET0 prediction in data-limited semi-arid regions.
• Provided a robust, data-efficient, and theoretically sound modeling framework that significantly advances data-driven ET0 modeling architecture and has high potential for sustainable water resources management applications.

Funding

Not explicitly mentioned in the provided text.

Citation

@article{Sabegh2025Enhancing,
  author = {Sabegh, Sanaz Monavvar and Zarehaghi, Davoud and Samadianfard, Saeed},
  title = {Enhancing reference evapotranspiration prediction with biological ensemble support vector regression and MODIS data integration},
  journal = {Sustainable Water Resources Management},
  year = {2025},
  doi = {10.1007/s40899-025-01317-1},
  url = {https://doi.org/10.1007/s40899-025-01317-1}
}