Athira et al. (2025) Multi-model ensemble improves evapotranspiration estimation over India

Identification

• Journal: Irrigation Science
• Year: 2025
• Date: 2025-11-28
• Authors: K V Athira, Eswar Rajasekaran, Gilles Boulet, Rahul Nigam, Bimal K. Bhattacharya
• DOI: 10.1007/s00271-025-01062-5

Research Groups

• Department of Civil Engineering, Indian Institute of Technology Bombay, Mumbai, India
• Centre for Climate Studies, Indian Institute of Technology Bombay, Mumbai, India
• CESBIO, Université de Toulouse, CNES, CNRS, INRAE, Toulouse, France
• Indo-French Cell for Water Science, ICWaR Indian Institute of Science, Bangalore, India
• Space Applications Centre (SAC, ISRO), Ahmedabad, India

Short Summary

This study develops and evaluates multi-model ensemble techniques for improving evapotranspiration (ET) estimation over India at both in-situ and 1 km spatial scales, demonstrating that machine learning-based ensembles significantly enhance accuracy compared to individual models and simpler ensemble methods.

Objective

• To develop and evaluate ensemble evapotranspiration (ET) models over the Indian region to improve the accuracy and reliability of ET estimation, addressing the spatiotemporal variability challenges of individual remote sensing-based models.

Study Configuration

• Spatial Scale: Seven in-situ sites across India (Nawagam, Samastipur, Jaisalmer, Malegaon, Berambadi, Dharwad, Kanpur) and country-wide coverage of India at 1 km spatial resolution.
• Temporal Scale: Daily ET for in-situ scale; 8-day periods for 1 km scale.

Methodology and Data

• Models used:
  • Individual ET models: Priestley-Taylor Jet Propulsion Lab (PT-JPL), Soil Plant Atmosphere and Remote Sensing Evapotranspiration (SPARSE – Layer and Patch), Surface Temperature Initiated Closure (STIC).
  • Ensemble techniques: Simple Averaging (Mean), Bayesian Model Averaging (BMA), k-Nearest Neighbor (k-NN), Random Forest (RF), Support Vector Machine (SVM).
• Data sources:
  • In-situ: Eddy Covariance (EC) flux towers and meteorological stations for daily ET validation.
  • Satellite: Moderate Resolution Imaging Spectroradiometer (MODIS) for Land Surface Temperature (LST), Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and albedo (MOD11A1/MYD11A1, MOD13Q1/MYD13Q1, MCD43A3 products). Landsat 8 for NDVI at in-situ sites.
  • Reanalysis: Indian Monsoon Data Assimilation and Analysis (IMDAA) for meteorological variables (12 km resolution, hourly).
  • Ancillary: Crop height measurements and empirical estimations.

Main Results

• Ensemble models consistently outperformed individual models at both in-situ and 1 km scales.
• At the field scale, individual ET model Root Mean Square Error (RMSE) ranged from 60 W m⁻² to 88 W m⁻², while ensemble RMSE ranged from 36 W m⁻² to 52 W m⁻².
• At the 1 km scale using MODIS data, individual model RMSE ranged from 46 W m⁻² to 52 W m⁻², and ensemble RMSE ranged from 26 W m⁻² to 39 W m⁻².
• Machine learning (ML) based ensembles (k-NN, RF, SVM) significantly reduced RMSE compared to simple averaging and BMA at both scales, with nearly half the error reduction at the field scale.
• Leave-one-out cross-validation demonstrated the spatial extrapolation capability of ensemble models, showing better performance than individual models at most sites.
• Land-cover-specific training further improved ET estimates, suggesting enhanced accuracy when sufficient data is available for specific land cover types.
• The ensemble models effectively captured the spatial variability of ET across India.

Contributions

• This study is the first to explore spatial-scale ensemble ET estimation over India using multiple remote sensing-based ET models.
• It demonstrates the superior performance of machine learning-based ensemble techniques (k-NN, RF, SVM) over traditional methods (simple mean, BMA) for ET estimation in the Indian region.
• The research validates the spatial extrapolation capability of ensemble models using leave-one-out cross-validation, highlighting their potential for regions with limited in-situ data.
• It identifies that land-cover-specific training can further enhance ET estimation accuracy when sufficient data is available.

Funding

• Space Applications Centre (ISRO), Ahmedabad (SHRESTI program, Project code: RD/0119-ISRO000-001)
• Geospatial Information Science and Engineering Hub at IIT Bombay (Project code: RD/0122-GISIR00-004)

Citation

@article{Athira2025Multimodel,
  author = {Athira, K V and Rajasekaran, Eswar and Boulet, Gilles and Nigam, Rahul and Bhattacharya, Bimal K.},
  title = {Multi-model ensemble improves evapotranspiration estimation over India},
  journal = {Irrigation Science},
  year = {2025},
  doi = {10.1007/s00271-025-01062-5},
  url = {https://doi.org/10.1007/s00271-025-01062-5}
}