Chen et al. (2025) An interpretable LAI time series prediction model of subtropic forests using a ConvLSTM coupling spatiotemporal attention mechanism model

Identification

• Journal: Ecological Indicators
• Year: 2025
• Date: 2025-11-19
• Authors: Li Chen, Xuejian Li, Huaqiang Du, Fangjie Mao, Hongyu Zhu, Jie Xuan, Yinyin Zhao, Zihao Huang, Kehan Mo, Yuanqing Zheng
• DOI: 10.1016/j.ecolind.2025.114437

Research Groups

• Key Laboratory of Carbon Sequestration and Emission Reduction in Agriculture and Forestry of Zhejiang Province, Zhejiang A&F University, Hangzhou, PR China
• College of Environment and Resources, College of Carbon Neutrality, Zhejiang A&F University, Hangzhou, PR China

Short Summary

This study developed an interpretable ConvLSTM model with a Spatio-Temporal Attention Mechanism (ConvLSTM-STAM) to predict forest Leaf Area Index (LAI) in subtropical forests of Zhejiang Province (2013–2018). The model achieved high accuracy (R² = 0.887, RMSE = 0.349 m²/m²) and provided mechanistic insights into seasonal LAI drivers through SHAP analysis.

Objective

• Propose and validate the ConvLSTM-STAM model for high-precision forest LAI time series prediction in subtropical forests.
• Quantify spatial and temporal contributions of various features to LAI prediction.
• Utilize SHAP analysis to interpret feature contributions, thereby enhancing ecological insights and practical applicability of the predictions.

Study Configuration

• Spatial Scale: Zhejiang Province, China. Data were processed at a 500 m spatial resolution.
• Temporal Scale: 2013–2018. Input data had temporal resolutions of 8 days (MODIS LAI), 16 days (MODIS NDVI, EVI), and daily (meteorological data), with ground observations collected monthly.

Methodology and Data

• Models used:
  • ConvLSTM-STAM (Convolutional Long Short-Term Memory with Spatio-Temporal Attention Mechanism) for LAI prediction.
  • Benchmark models: LSTM, LSTM-STAM, CNN-LSTM-STAM.
  • SHAP (SHapley Additive exPlanations) for model interpretability.
  • LACC (Locally Adjusted Cubic-spline Capping) algorithm for MODIS LAI smoothing.
  • Modified Savitzky-Golay (mSG) filter for NDVI and EVI smoothing.
  • Inverse Distance Weighting (IDW) interpolation for meteorological data downscaling.
• Data sources:
  • Remote Sensing: MODIS (MOD15A2H for LAI/FPAR, MOD13A1 for NDVI/EVI/NIR, MCD12Q1 for land cover types including evergreen needleleaf, evergreen broadleaf, deciduous broadleaf, and mixed forests) obtained from Google Earth Engine (GEE).
  • Meteorological: Daily maximum temperature (Tmax), minimum temperature (Tmin), precipitation (Pre), and total solar radiation (Rad) from the China Meteorological Science Data Sharing Service.
  • Topographic: Elevation (DEM), slope, and aspect from the Geospatial Data Cloud platform, Chinese Academy of Sciences.
  • Ground Observation: Monthly ground-based LAI data collected from permanent sample plots in Shanchuan Township, Taihuyuan Town, and Tianmushan of Zhejiang Province using a WinSCANOPY 2009a canopy imager.

Main Results

• The ConvLSTM-STAM model achieved superior LAI prediction accuracy (R² = 0.887, RMSE = 0.349 m²/m²), significantly outperforming MODIS_LAI (R² = 0.443, RMSE = 0.774 m²/m²) and benchmark deep learning models (improving R² by 15.8–42.4 % and reducing RMSE by 30.5–45.2 %).
• The model successfully captured the seasonal phenology of forest LAI and its spatial heterogeneity across Zhejiang Province, showing higher LAI (4–7.5 m²/m²) in southwestern regions and lower LAI (0–2 m²/m²) in northeastern and central areas.
• SHAP analysis identified LAI (autoregressive), minimum temperature (Tmin), average temperature (Tavg), maximum temperature (Tmax), and Normalized Difference Vegetation Index (NDVI) as the most influential features for LAI prediction.
• Seasonal SHAP analysis revealed dynamic feature contributions: LAI contributed positively in summer and autumn but negatively in spring and winter, while Tmin exhibited the opposite trend. Tmin and Tavg dominated LAI variability in spring and winter, with LAI itself being the leading factor in summer and autumn.
• Feature interaction analysis demonstrated that existing LAI levels modulate the effects of climatic and vegetation variables, with temperature and NDVI effects amplified at higher LAI levels.
• Spatio-temporal attention mechanism (STAM) weights showed spatially heterogeneous importance, with higher attention in eastern, southern, and northwestern Zhejiang, peaking in summer, consistent with seasonal vegetation growth patterns.

Contributions

• First application of a ConvLSTM-STAM model combined with SHAP analysis for interpretable LAI time series prediction in subtropical forests, providing a scalable and interpretable framework for ecological monitoring.
• Developed a novel interpretable deep learning framework that simultaneously enhances LAI prediction accuracy and provides mechanistic insights into spatiotemporal drivers of LAI.
• Integrated multi-source remote sensing, meteorological, and topographic data to effectively address challenges posed by spatial heterogeneity and temporal dynamics in complex subtropical forest ecosystems.
• Quantified individual and synergistic feature contributions to LAI prediction across space and time, revealing crucial seasonal shifts in the importance of climatic and vegetation drivers.
• Generated high-accuracy, interpretable LAI time series products for Zhejiang Province (2013–2018), offering valuable resources for forest carbon accounting, precision forestry, and ecosystem management in similar subtropical ecosystems globally.

Funding

• National Natural Science Foundation of China (No. 32201553, 32171785)
• Leading Goose Project of Science Technology Department of Zhejiang Province (No. 2023C02035)
• Scientific Research Project of Baishanzu National Park (No. 2022JBGS02)
• Natural Science Fund of Huzhou (2022YZ33)

Citation

@article{Chen2025interpretable,
  author = {Chen, Li and Li, Xuejian and Du, Huaqiang and Mao, Fangjie and Zhu, Hongyu and Xuan, Jie and Zhao, Yinyin and Huang, Zihao and Mo, Kehan and Zheng, Yuanqing},
  title = {An interpretable LAI time series prediction model of subtropic forests using a ConvLSTM coupling spatiotemporal attention mechanism model},
  journal = {Ecological Indicators},
  year = {2025},
  doi = {10.1016/j.ecolind.2025.114437},
  url = {https://doi.org/10.1016/j.ecolind.2025.114437}
}