Interpretable machine-learning diagnosis of forest gross primary productivity patterns in China’s protected areas

Understanding the spatial variability of forest gross primary productivity (GPP) is essential for diagnosing ecosystem functioning and supporting monitoring and management within protected areas. However, many data-driven approaches emphasize predictive performance while offering limited interpretability of the drivers underlying spatial heterogeneity. Here, we develop an interpretable machine-learning framework to diagnose spatial patterns and dominant drivers of forest GPP within China’s national-level protected areas. Satellite-derived GPP and environmental variables were aggregated to a 0.1° spatial grid over 1990–2018 to characterize long-term mean forest productivity. Multiple machine-learning models were evaluated, and the best-performing model was interpreted using explainable artificial intelligence to quantify driver importance and response behavior. The mean forest GPP was 759.5 g C m⁻2 yr⁻1, with pronounced spatial heterogeneity. Approximately 22% of protected areas experienced increases in forest GPP exceeding 20% between 1990 and 2018, primarily in humid regions, whereas 12.5% showed declines greater than 20%, mainly in the southern Qinghai-Tibet Plateau, northern arid regions, and west-central temperate semihumid zones. Among the evaluated models, XGBoost achieved the highest predictive performance on independent test data (R2 = 0.76, RMSE = 262 g C m⁻2 yr⁻1). Precipitation, temperature, and solar radiation emerged as the dominant drivers, with precipitation explaining 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. This study is a spatial diagnostic analysis that provides transparent insights to support spatial prioritization, monitoring design, and management planning within protected areas.