Abstract
Net ecosystem productivity (NEP) in croplands is a core indicator of carbon exchange between agroecosystems and the atmosphere, directly reflecting net carbon budgets and sequestration capacity. To resolve its spatiotemporal patterns and dominant controls, we combined remote-sensing, land-cover, and meteorological datasets with the Boreal Ecosystem Productivity Simulator (BEPS) coupled to a Geostatistical Model of Soil Respiration (GSMSR) to simulate cropland NEP across China's three major plains-the Northeast China Plain (NCP), Huang-Huai-Hai Plain (HHHP), and Middle-Lower Yangtze Plain (MLYP)-during 2000-2020. An interpretable machine-learning framework (XGBoost-SHAP) was used to quantify factor responses and regional heterogeneity. The results show that: (1) From 2000 to 2020, cropland NEP across the three major plains increased overall but exhibited a pronounced north-south gradient: cropland NEP increases were larger and more widespread in NCP and HHHP, while persistent negative changes were concentrated in southern MLYP. (2) Analysis of influencing factors revealed that interannual variations in agricultural NEP from 2000 to 2020 were jointly regulated by hydro temperature factors, atmospheric composition, and soil and farm management factors. The important ranking of factors varies with regional heterogeneity. (3) Regionally, NCP was primarily driven by annual mean temperature and surface soil moisture, HHHP was dominated by annual mean temperature and carbon dioxide, while MLYP was influenced mainly by annual mean temperature and PM(2.5). Threshold effects were observed for all factors. Notably, declining PM(2.5) concentrations exerted a positive influence on interannual variations in cropland NEP. This study can provide scientific basis for safeguarding food security and advancing sustainable agricultural development and offer reference for formulating cross-regional policies on enhancing carbon sequestration in croplands and implementing zoned management.