Cabral et al. (2026) Interpretable machine-learning diagnosis of forest gross primary productivity patterns in China’s protected areas

Identification

• Journal: International Journal of Applied Earth Observation and Geoinformation
• Year: 2026
• Date: 2026-04-02
• Authors: Pedro Cabral, Xiaofeng Ren, Chenxi Zhu, Emmanuel Yeboah, Guojie Wang, Erwen Xu, Wenmao JING, Alberto Bento Charrua, Oualid Hakam, Ana Cláudia Coimbra Costa
• DOI: 10.1016/j.jag.2026.105270

Research Groups

• School of Remote Sensing and Geomatics Engineering, Nanjing University of Information Science & Technology, Nanjing, China
• Gansu Qilian Mountain Water Conservation Forest Research Institute, Zhangye, China
• Qilian Mountain Eco-Environment Research Center of Gansu Province, Lanzhou, China
• Gansu Qilian Mountain Forest Ecosystem of the State Research Station, Zhangye, Gansu, China
• Department of Earth Sciences and Environment, Faculty of Science and Technology, Licungo University, Beira, Mozambique
• Center for Remote Sensing Applications, Mohammed VI Polytechnic University, Ben-Guerir, Morocco
• NOVA Information Management School (NOVA IMS), Universidade Nova de Lisboa, Lisboa, Portugal

Short Summary

This study developed an interpretable machine-learning framework to diagnose spatial patterns and dominant drivers of forest gross primary productivity (GPP) in China's national-level protected areas, finding that precipitation, temperature, and solar radiation are the primary drivers, with precipitation being the most dominant factor across the study area.

Objective

• To develop an interpretable machine-learning framework to diagnose spatial patterns and dominant drivers of forest gross primary productivity (GPP) within China’s national-level protected areas from 1990 to 2018, enhancing understanding of driver importance and response behavior.

Study Configuration

• Spatial Scale: Forested national-level protected areas across mainland China, analyzed on a 0.1-degree spatial grid (53,314 spatial sampling units).
• Temporal Scale: 29-year period (1990–2018), with satellite-derived GPP aggregated to annual totals and all variables represented by long-term mean values for spatial analysis.

Methodology and Data

• Models used:
  • Machine Learning: Extreme Gradient Boosting (XGBoost, selected as best-performing), Random Forest, Linear Regression, Classification and Regression Tree (CART).
  • Explainable Artificial Intelligence (XAI): SHapley Additive exPlanations (SHAP), Partial Dependence Plots (PDPs).
  • Causal Inference: Convergent Cross Mapping (CCM), Geographical Convergent Cross Mapping (GCCM).
  • Trend Analysis: Theil-Sen median slope estimator, Mann-Kendall (MK) test.
  • Multicollinearity Analysis: Variance Inflation Factor (VIF), Tolerance (TOL).
• Data sources:
  • Forest Gross Primary Productivity (GPP): Satellite-driven NIRv-based GPP dataset for China (S. Wang et al., 2021; Zhu et al., 2025).
  • Climatic variables: Gridded climate datasets for precipitation, temperature (Peng et al., 2019), and solar radiation (He et al., 2020).
  • Environmental and anthropogenic drivers: Soil moisture (Zhang et al., 2024a), Nighttime light (Zhang et al., 2024b), Forest fragmentation index (Yang and Huang, 2021), Elevation, Aspect, Slope (CGIAR-CSI, 2022), Climate zones (Kottek et al., 2006).
  • Protected area boundaries: Chinese Nature Reserve Specimen Resource Sharing Platform.

Main Results

• The mean forest GPP within China's protected areas from 1990 to 2018 was 759.5 g C m⁻² yr⁻¹.
• Forest GPP increased at an average rate of 1.8703 g C m⁻² yr⁻¹ (P < 0.05) during the study period.
• Spatial heterogeneity in GPP change was pronounced: approximately 22% of protected areas experienced increases exceeding 20% (mainly in humid regions), while 12.5% showed declines greater than 20% (southern Qinghai-Tibet Plateau, northern arid regions, west-central temperate semihumid zones).
• XGBoost achieved the highest predictive performance (R² = 0.76, RMSE = 262 g C m⁻² yr⁻¹).
• The dominant drivers of forest GPP, ranked by mean absolute SHAP value, were precipitation (157.63), temperature (99.81), solar radiation (79.26), and forest fragmentation (52.22).
• Spatially, precipitation was the dominant driver in 53.4% of the study area, followed by temperature (19.7%) and solar radiation (16.0%).
• Forest fragmentation exhibited a predominantly negative association with forest GPP.
• Partial Dependence Plots indicated highest GPP values when temperature exceeded approximately 18 °C, solar radiation ranged between 720 and 736 W m⁻², and precipitation ranged between 100 and 360 mm.
• Convergent Cross Mapping (CCM) revealed statistically significant directional associations (P < 0.001) from all examined drivers to forest GPP, with no significant reverse directional associations detected.

Contributions

• Integrates explainable artificial intelligence (XAI) and spatiotemporal causal analysis to provide transparent, spatially explicit insights into the dominant drivers and their interactions shaping forest GPP patterns in China's protected areas.
• Moves beyond traditional predictive performance by quantifying driver importance, visualizing response behaviors, and identifying directional associations, enhancing ecological interpretation.
• Offers a diagnostic analysis to support spatial prioritization, monitoring design, and management planning within protected areas, highlighting climate-limited regions and areas influenced by human activities.
• Provides a comprehensive long-term (1990-2018) assessment of forest GPP dynamics and its drivers across diverse climatic zones within China's protected areas.

Funding

• National Natural Science Foundation of China (#42275028, U22A20592)
• Gansu Province Science and Technology Plan Project (24JRRG034, 25JRRG027), China
• Gansu Province International Science and Technology Cooperation Project (25YFWG001), China
• The Special Project for Improving Scientific Research Conditions under the Science and Technology Plan of Zhangye City (ZY2024KY01), China
• FCT (Fundação para a Ciência e a Tecnologia), under the project - UID/04152/2025 - Centro de Investigação em Gestão de Informação (MagIC)/NOVA IMS - Doi: 10.54499/UID/04152/2025 (2025-01-01/2028-12-31) and UID/PRR/04152/2025 Doi: 10.54499/UID/PRR/04152/2025 (2025-01-01/2026-06-30)

Citation

@article{Cabral2026Interpretable,
  author = {Cabral, Pedro and Ren, Xiaofeng and Zhu, Chenxi and Yeboah, Emmanuel and Wang, Guojie and Xu, Erwen and JING, Wenmao and Charrua, Alberto Bento and Hakam, Oualid and Costa, Ana Cláudia Coimbra},
  title = {Interpretable machine-learning diagnosis of forest gross primary productivity patterns in China’s protected areas},
  journal = {International Journal of Applied Earth Observation and Geoinformation},
  year = {2026},
  doi = {10.1016/j.jag.2026.105270},
  url = {https://doi.org/10.1016/j.jag.2026.105270}
}