Answers are grounded in this paper and cite their sources.
Climate uncertainty threatens agricultural productivity and requires adaptive crop management strategies. Planting date is a key determinant of yield, yet its optimization remains challenging due to nonlinear interactions with climate variability and spatial heterogeneity. This study develops an integrated framework combining the process-based APEX model, machine learning (ML) emulation, and reinforcement learning (RL) to explore adaptive planting policies under historical and future climate scenarios (RCP 4.5 and RCP 8.5). Corn production was simulated for 203 fields in Illinois using APEX driven by downscaled climate projections. The planting window (April 5–May 17) was discretized into six weekly options defining contextual bandit action space. An ML emulator was trained on APEX outputs to approximate yield responses to climatic, spatial, and management variables, with XGBoost achieving the highest accuracy. This emulator served as the reward function within a contextual bandit RL framework, where each season was treated as an independent decision episode. APEX results show a decline in mean yield from ∼10 to <6 Mg ha⁻¹ by the late century, especially under RCP 8.5, driven by increasing heat stress and climate variability. The RL agent learned effective policies but exhibited asymptotic partial convergence, reflecting an attenuated reward signal from planting date relative to dominant environmental drivers. The highest-average planting policies shifted from early planting under historical conditions to later or intermediate dates under future climates. However, re-simulations using a single representative regional planting policy produced only modest yield gains, reflecting strong field-level heterogeneity and the existence of multiple near-optimal planting actions. These findings emphasize the need for spatially adaptive decision frameworks, demonstrating that the proposed APEX–ML–RL architecture provides a scalable simulation-based framework for investigating climate-adaptive agricultural policies and a flexible foundation for future site-specific crop management applications.
No takes yet. Share an insight, caveat, or question.
Noa-Yarasca et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: