Sha et al. (2026) Simulating ecosystem water and carbon fluxes by integrating remote sensing indices with multiple machine learning models

Identification

• Journal: Agricultural Water Management
• Year: 2026
• Date: 2026-07-29
• Authors: Lixuan Sha, Zhifang Feng, Ting Zhang, Wenjie Quan, Keke Zhou, Jianzhu Li
• DOI: 10.1016/j.agwat.2026.110622

Research Groups

• State Key Laboratory of Hydraulic Engineering Intelligent Construction and Operation, Tianjin University, China
• China Communications Construction Company Tianjin Dredging Co., Ltd., China
• PowerChina Huadong Engineering Corporation Limited, China

Short Summary

This study integrated remote sensing indices with multiple machine learning models to simulate ecosystem water and carbon fluxes. The results showed that incorporating remote sensing data significantly improved the simulation accuracy for carbon fluxes, particularly gross primary production (GPP) and net ecosystem exchange (NEE).

Objective

• Evaluate model performance across four key land-atmosphere fluxes: GPP, evapotranspiration (ET), NEE, and ecosystem respiration (Reco)
• Investigate the potential of interpretable machine learning in disentangling non-linear roles of functional and structural vegetation indices

Study Configuration

• Spatial Scale: Point-scale observation region centered around the Liulin Eco-hydrological Observation Station (114◦21'E, 37◦17'N) with a catchment area of 57.4 km2.
• Temporal Scale: Daily time step for flux simulations during the 2021–2023 growing seasons.

Methodology and Data

• Models used: Five machine learning algorithms: Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Categorical Boosting (CatBoost), and Natural Gradient Boosting (NGBoost).
• Data sources: In-situ flux and hydrometeorological observations, satellite-derived solar-induced chlorophyll fluorescence (SIF) and leaf area index (LAI) products.

Main Results

• Incorporating remote sensing data significantly improved the simulation accuracy for GPP and NEE.
• The CatBoost algorithm outperformed other configurations for GPP and Reco simulations, while XGBoost was selected for ET and NEE.
• SHAP analysis revealed that SIF, LAI, and PAR were the predominant predictive drivers for GPP, contributing mean absolute SHAP values of approximately 0.80, 0.70, and 0.50 g C m−2 d−1, respectively.

Contributions

• This study demonstrates the potential of interpretable machine learning in disentangling non-linear roles of functional and structural vegetation indices.
• The derived quantitative benchmarks provide a rigorous, data-driven framework for evaluating site-specific ecosystem responses to hydroclimatic stressors.

Funding

• Not specified

Citation

@article{Sha2026Simulating,
  author = {Sha, Lixuan and Feng, Zhifang and Zhang, Ting and Quan, Wenjie and Zhou, Keke and Li, Jianzhu},
  title = {Simulating ecosystem water and carbon fluxes by integrating remote sensing indices with multiple machine learning models},
  journal = {Agricultural Water Management},
  year = {2026},
  doi = {10.1016/j.agwat.2026.110622},
  url = {https://doi.org/10.1016/j.agwat.2026.110622}
}