We use AI to build China's first county-level agricultural policy database and identify which interventions effectively protect crop yields from climate shocks.
Agricultural productivity on the North China Plain underpins national food security, yet climate hazards increasingly threaten the winter wheat–summer maize system that dominates this region. County governments implement diverse policies targeting specific production stages, from fertilization and irrigation to residue management. Whether these policies effectively buffer productivity against climate shocks has not been systematically assessed, largely because no structured record of county-level policy interventions exists. Here, we construct such a database by extracting policy documents from the open government portals of over 300 counties using large language models. Each document is parsed into structured fields: the production stage targeted, the policy instrument deployed, and the intervention intensity. We then apply machine-learning-augmented breakpoint detection to county-level yield time series to identify where and when productivity underwent significant shifts, and match detected breakpoints to the policy timeline. For each policy–breakpoint pair, we trace the causal pathway through a crop production cycle framework that maps how interventions alter exposure or sensitivity to climate hazards at each stage. Pilot results from Shandong Province indicate that policies can be reliably extracted and classified, that breakpoints cluster around known policy rollouts, and that the production cycle framework distinguishes policies raising average yield from those reducing climate-induced yield variability.
I am a researcher at Nanjing University, China, specializing in land-use policy evaluation using causal inference methods. My work links agricultural investment to food security and biodiversity