Key points are not available for this paper at this time.
Understanding spatially heterogeneous controls on cropland productivity is critical for improving agroecosystem sustainability under increasing climatic variability and anthropogenic pressure. However, the nonlinear and spatially non-stationary relationships between productivity and its drivers remain difficult to quantify using conventional approaches. Using multi-source remote sensing and ancillary datasets, this study applied an interpretable machine-learning approach to investigate the spatial controls of cropland net primary productivity (NPP) across humid subtropical China based on two representative periods (2010 and 2020). Results showed that cropland NPP increased by 4.7% during the study period, although substantial spatial heterogeneity persisted. Climatic factors, particularly potential evapotranspiration and mean annual temperature, showed the strongest contributions to spatial variations in NPP, with their contribution patterns exhibiting pronounced spatial reversals along hydroclimatic gradients. Soil pH acted as a conditional regulator within a narrow optimal range (5.5–6.0), whereas nitrogen fertilizer input and population density imposed increasingly negative effects beyond critical thresholds. Clustering of pixel-level SHAP vectors identified four functional zones characterized by distinct combinations of dominant drivers and constraint regimes, linking regional productivity gains with localized vulnerability risks. The results further reveal that cropland productivity exhibits multiple interacting constraint domains, where productivity responses depend on whether local conditions remain within effective threshold ranges. By integrating remote sensing observations with explainable machine learning, this study provides a spatially explicit framework for understanding productivity–ecological environment relationships and supporting adaptive cropland management under climate change.
Liao et al. (Thu,) studied this question.