From rules to learning: A GIS-integrated agent-based reinforcement learning framework for simulating farmers’ land-use decisions and sustainable policy design

Developing countries face critical tensions between farmland-use policies prioritizing food security and environmental sustainability, and farmers’ livelihood-centered land-use strategies. To address these challenges, this study develops a human-centered agent-based model that integrates dual-layer Q-learning with GIS-based spatial representation. This framework simulates farmers’ adaptive decision-making under dynamic policy and environmental conditions while formalizing behavioral heterogeneity through Theory of Planned Behavior (TPB)-informed typologies. Agents dynamically evolve land-transfer and cropping strategies via perception–action–feedback learning cycles. Applied to and validated against empirical observations from a representative Chinese case, the model reveals non-linear scale–efficiency relationships, with medium-scale farms exhibiting the comparatively balanced performance across yield and multiple resource-use efficiency dimensions. The analysis further identifies evolutionary pathways of adaptive behavior under policy incentives and quantifies trade-offs among area-weighted crop output, resource-use efficiency, and regional gross income. Scenario simulations show that rice-oriented subsidies increase rice output but reduce resource-use efficiency, rapeseed output, and gross income, whereas rotation-oriented incentives increase gross income while maintaining rice output near the baseline level. These findings support performance-linked subsidies based on multidimensional outcomes, tiered land-transfer regulation, and diversified cropping incentives. By linking adaptive behavioral dynamics with spatial and institutional constraints, this study provides a robust analytical framework for evaluating land-use policies and supports more effective governance strategies for balancing food security, environmental sustainability, and rural livelihoods.