Zhou et al. (2026) Optimizing light gradient boosting machine with the slime mould algorithm for reference evapotranspiration estimation

Identification

• Journal: Agricultural Water Management
• Year: 2026
• Date: 2026-01-06
• Authors: Hanmi Zhou, Yumin Su, Linshuang Ma, J LI, Sibo Lu, Cheng Chen, Youzhen Xiang, Rui Li, Zhe Peng, Ru Huang
• DOI: 10.1016/j.agwat.2025.110107

Research Groups

• College of Agricultural Equipment Engineering, Henan University of Science and Technology, Luoyang, Henan Province, China
• Key Laboratory of Agricultural Soil and Water Engineering in Arid and Semiarid Areas of Ministry of Education, Northwest A&F University, Xianyang, Shaanxi Province, China

Short Summary

This study developed a novel hybrid SMA-LGB model for accurate reference evapotranspiration (ETo) estimation in the Songliao Plain, demonstrating superior accuracy and robustness compared to traditional models, even with limited or missing meteorological data.

Objective

• Identify key meteorological factors affecting ETo in the Songliao Plain through feature importance analysis, providing a basis for ETo estimation under data-limited conditions.
• Construct the SMA-LGB model and compare its performance with the original LGB model and other models to validate its advantages.
• Evaluate the portability of the SMA-LGB model through cross-site validation, ensuring ETo estimation for target sites under data scarcity.
• Provide scientific support for regional irrigation scheduling optimization and efficient agricultural water resource utilization.

Study Configuration

• Spatial Scale: Songliao Plain, Northeast China (40°–48°N, 120°–127°E), using data from 25 representative meteorological stations.
• Temporal Scale: Daily meteorological observation records spanning 30 years, from January 1990 to December 2019, totaling 10,957 data points per station.

Methodology and Data

• Models used:
  • Hybrid Optimization Model: Slime Mould Algorithm-Light Gradient Boosting (SMA-LGB)
  • Machine Learning Models: Light Gradient Boosting (LGB), Extreme Gradient Boosting (XGB), Adaptive Boosting (ADA), k-Nearest Neighbors (KNN)
  • Reference Model: FAO-56 Penman-Monteith (FAO-56 PM)
  • Empirical Model: Hargreaves-Samani (H-S)
  • Feature Selection Method: Gradient Boosting Decision Tree (GBDT)
• Data sources:
  • Daily meteorological observation records from the National Meteorological Information Center of the China Meteorological Administration (CMA).
  • Variables: daily minimum temperature (TMN), maximum temperature (TMX), relative humidity (RHU), sunshine duration (SSD), and wind speed at 2 meters (U2).
  • Derived data: Extraterrestrial radiation (Ra) calculated based on site latitude and day of the year.
  • Data preprocessing: Quality control, outlier identification (IQR method), and missing value imputation (K-nearest neighbors method).

Main Results

• Meteorological Factor Importance: The most influential factors for ETo estimation were TMX (mean importance score = 0.429) and Ra (mean importance score = 0.351), with TMX dominant in central-northern regions and Ra in southern regions. Subsequent factors were SSD (0.096), RHU (0.072), and U2 (0.039).
• SMA-LGB Model Performance:
  • The SMA-LGB model consistently demonstrated superior accuracy and stability compared to LGB, XGB, ADA, and KNN across all input combinations (M1: TMX, Ra, SSD; M2: TMX, Ra, SSD, RHU; M3: TMX, Ra, SSD, RHU, U2).
  • For the M3 combination, SMA-LGB achieved average R² of 0.998, NSE of 0.998, RMSE of 9.6 × 10⁻⁵ m⋅d⁻¹, and MAE of 6.7 × 10⁻⁵ m⋅d⁻¹.
  • SMA-LGB reduced the average RMSE by 16.5% compared to classical models across the three input combinations.
  • The model exhibited superior generalization capability with a small performance gap between training and testing datasets.
• Spatial Variability of Accuracy: ETo estimation accuracy improved from northern to southeastern stations, with RMSE values decreasing across this gradient. The SMA-LGB model consistently achieved the lowest RMSE values across all regions.
• Cross-site Portability: When trained with data from adjacent meteorological stations, the SMA-LGB model maintained high ETo estimation accuracy (R² > 0.95), even when data were missing at the target station. Integrating data from multiple neighboring stations further improved accuracy, reducing average RMSE by 4.7% compared to single-site training.
• Comparison with Hargreaves-Samani Model: The SMA-LGB model (with TMX, TMN, Ra inputs) outperformed the locally calibrated H-S model, reducing RMSE by an average of 11.7%. When driven by the three key factors (TMX, Ra, SSD), SMA-LGB achieved even better performance, increasing R² by 7.5% and decreasing RMSE by 26.8% on average compared to the H-S model.

Contributions

• First application of the hybrid Slime Mould Algorithm-Light Gradient Boosting (SMA-LGB) framework in the agricultural water resources domain for reference evapotranspiration (ETo) estimation.
• Quantified the relative contributions of various meteorological factors to ETo estimation in the Songliao Plain using a GBDT-based feature selection method.
• Developed a novel hybrid SMA-LGB model that significantly enhances both the accuracy and robustness of ETo estimation, outperforming traditional machine learning models and the locally calibrated Hargreaves-Samani empirical model.
• Demonstrated the strong cross-site adaptability and reliability of the SMA-LGB model for ETo estimation under conditions of limited or missing meteorological data at target stations.
• Provided a reliable method for ETo estimation under data-scarce conditions, offering scientific support for precision irrigation and refined agricultural water resource management in the Songliao Plain and similar climatic regions.

Funding

• National Natural Science Foundation of China (52379039, 51909079)
• Program for Science & Technology Innovation Talents in Universities of Henan Province (25HASTIT012)

Citation

@article{Zhou2026Optimizing,
  author = {Zhou, Hanmi and Su, Yumin and Ma, Linshuang and LI, J and Lu, Sibo and Chen, Cheng and Xiang, Youzhen and Li, Rui and Peng, Zhe and Huang, Ru},
  title = {Optimizing light gradient boosting machine with the slime mould algorithm for reference evapotranspiration estimation},
  journal = {Agricultural Water Management},
  year = {2026},
  doi = {10.1016/j.agwat.2025.110107},
  url = {https://doi.org/10.1016/j.agwat.2025.110107}
}