An LLM-driven agent prototype for automated urban expansion simulation

To address fragmented toolchains and the high costs of hyper-parameter tuning and reproducibility in urban-expansion simulation, this paper proposes and prototypes LUNA-Agent, an LLM-driven system. The agent uses natural language as a unified interface to organize factor selection, model training and inference, accuracy evaluation, visualization, and parameter-level reflection into an auditable closed loop. Without modifying any training or inference code, it conducts minimal adjustments, including class weights, negative/positive sampling ratio, and thresholding policy, to enable automatic tuning. In a 2015–2020 case study of Lanzhou with three factors (nighttime lights, log distance-to-road, and slope), the system reliably runs Random Forest (RF), Histogram-based Gradient Boosting Trees (HGBT), and Logistic Regression (LR) and achieves overall accuracy (OA) ≈ 95.8% and Cohen’s Kappa ≈ 0.85. With the reflection mechanism enabled, Kappa, F1, and the map-level figure of merit (FoM) increase monotonically across iterations, and FoM improves by approximately 47%, indicating effectiveness for sparse-change detection. This prototype-level validation performs data acquisition and preprocessing manually and uses only three factors. The results confirm the feasibility and promise of LLM agents for urban-expansion simulation and lay the groundwork for end-to-end, closed-loop automation.