Liu et al. (2025) Unraveling the impacts of climate factors on leaf area index of Chinese grasslands using interpretable machine learning models

Identification

• Journal: Theoretical and Applied Climatology
• Year: 2025
• Date: 2025-12-18
• Authors: Wei Liu, Chengxi Gao, Shaozhi Lin, Yu Zhou, Wenrui Bai, Junhu Dai, Huanjiong Wang
• DOI: 10.1007/s00704-025-05970-6

Research Groups

• Key Laboratory of Land Surface Pattern and Simulation, Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences
• University of Chinese Academy of Sciences

Short Summary

This study utilized three interpretable machine learning models to investigate the impact mechanisms of preseason climate and extreme weather events on grassland Leaf Area Index (LAI) in China from 2001 to 2020, finding that preseason climate was the most important driver, with extreme events and CO2 fertilization also significantly influencing LAI dynamics.

Objective

• To determine if machine learning models can accurately simulate the dynamic changes of grassland LAI.
• To investigate how climatic factors, such as preseason climate and extreme weather events, drive changes in grassland LAI.
• To assess differences in machine learning model performance in simulating grassland LAI across different grassland types and seasons.

Study Configuration

• Spatial Scale: Chinese grasslands, specifically 8 major grassland types in northwestern China, at a spatial resolution of 0.25 degrees.
• Temporal Scale: From 2001 to 2020, with LAI data at an 8-day resolution, climate data at hourly resolution, and CO2 concentration at monthly resolution.

Methodology and Data

• Models used:
  • Bayesian spatiotemporal mixed model (Bayes)
  • Interpretable neural network model based on generalized additive models with structured interactions (GAMI)
  • Random Forest model (RF)
• Data sources:
  • LAI: GLASS LAI Dataset (2001-2020, 0.05° upscaled to 0.25°, 8-day resolution).
  • Climate: ERA5 hourly reanalysis dataset (2001-2020, 0.25° resolution). Variables included 2 m temperature, 2 m dewpoint temperature, total precipitation, and surface net solar radiation, with derived variables for preseason periods (15, 30, 45, 60 days) such as daily mean temperature, vapor pressure deficit, net solar radiation, precipitation, and extreme event days (heat, frost, heavy-rainfall, no-rainfall).
  • Land Cover/Vegetation Type: China Land Cover Dataset (CLCD, 2001-2020, 30 m upscaled to 0.25°), 1:1 million Vegetation Dataset in China (1 km upscaled to 0.25°).
  • Soil: Harmonized World Soil Database (30 arcseconds upscaled to 0.25°). Variables included clay content, pH, organic carbon content, total nitrogen content, and carbon-to-nitrogen ratio.
  • CO2 concentration: Global Atmospheric Carbon Dioxide Concentration Simulation Grid Dataset (monthly, 2° × 2.5° grid resolution).
  • Grazing intensity: Long-term High-resolution Dataset of Grasslands Grazing Intensity in China (2001-2020, 0.1° resolution, in standard sheep units per hectare).

Main Results

• All three machine learning models performed well in simulating LAI (R² ranging from 0.540 to 0.963), with the Random Forest model demonstrating the best performance (R² = 0.963, RMSE = 0.098).
• Preseason climate was identified as the most important factor driving LAI changes across all models.
• Increases in preseason temperature, precipitation, and radiation generally led to an increase in grassland LAI.
• Extreme weather events showed significant impacts: heat events and heavy-rainfall events had positive effects on LAI, while frost events and no-rainfall events had negative impacts.
• CO2 concentration exhibited a significant fertilization effect on LAI.
• Grazing intensity had a relatively small and complex impact on LAI, showing a decrease when intensity was less than 1 standard sheep unit per hectare (SU/ha) and an increase when greater than 1 SU/ha.
• Model performance varied by grassland type and season; models performed better in meadow grasslands (higher precipitation) than steppe grasslands and better in spring than in autumn.
• The impact of precipitation on LAI was greater in spring than in autumn, whereas the impacts of temperature and radiation were greater in autumn than in spring.

Contributions

• Developed and validated a climate change-adaptive framework for predicting grassland growth dynamics using interpretable machine learning models.
• Provided a comprehensive investigation into the impact mechanisms of various climate factors, including preseason climate and extreme weather events, on grassland LAI in China.
• Offered a comparative analysis of the performance and interpretability of Bayesian, interpretable neural network, and Random Forest models for LAI simulation.
• Revealed the differential importance of driving factors and model performance across distinct grassland types and seasons, enhancing understanding of regional and seasonal LAI dynamics.
• Provided scientific support for grassland ecological protection and adaptive management strategies in the context of climate change.

Funding

• National Key Research and Development Program of China (grant no. 2023YFF1303804)
• National Natural Science Foundation of China (grant no. 42271062)

Citation

@article{Liu2025Unraveling,
  author = {Liu, Wei and Gao, Chengxi and Lin, Shaozhi and Zhou, Yu and Bai, Wenrui and Dai, Junhu and Wang, Huanjiong},
  title = {Unraveling the impacts of climate factors on leaf area index of Chinese grasslands using interpretable machine learning models},
  journal = {Theoretical and Applied Climatology},
  year = {2025},
  doi = {10.1007/s00704-025-05970-6},
  url = {https://doi.org/10.1007/s00704-025-05970-6}
}