A Coupled Cellular Automata-Deep Learning-XGBoost for Land Use and Land Cover Change Prediction in Yancheng

Accurate simulation of land use and land cover change (LUCC) is essential for understanding urban expansion and ecosystem evolution. However, existing models struggle to simultaneously capture complex spatiotemporal dependencies and nonlinear socioeconomic drivers, often failing when historical trends are disrupted by sudden policy interventions (i.e., structural breaks). To address these limitations, this study proposes a parallel mechanism-coupled Cellular Automata–Deep Learning–XGBoost (CDX) framework that dynamically integrates spatial, temporal, and socioeconomic driving forces. In this framework, a decoupled CNN-LSTM module extracts spatiotemporal inertia from remote sensing sequences; an XGBoost module transparently quantifies driver suitability; and a Cellular Automata (CA) module simulates spatial diffusion. These components are integrated through a unified transition potential function, optimized by adaptive stochastic search, to balance spatial realism with mechanism fidelity. The CDX framework was applied to Yancheng, China, using multi-temporal Landsat imagery (1995–2025). Experimental results validate the framework's superior mechanism adaptability: while purely data driven baselines (e.g., MicroViT) remained locked in historical inertia, the adaptively calibrated CDX successfully characterized the policy-driven regime shift from inertia-driven expansion to eco-regulated growth (2020–2025). Consequently, CDX achieved a diagnostic Figure of Merit (FoM) improvement of approximately 12.6% compared to the state-of-the-art MicroViT model and over 50% compared to the traditional RF-CA baseline. Leveraging the transition rules explicitly learned from this structural evolution, the model generated robust projections for the 2025–2050 period. Comparative scenario simulations reveal that while an economic priority scenario results in coastal sprawl, the baseline projection identifies an optimized trajectory where steady built-up land expansion (reaching 37.32% by 2050) is successfully decoupled from ecological degradation, enabling significant wetland recovery (increasing to 7.15%). This study demonstrates that the CDX framework effectively bridges spatial pattern simulation with the adaptive learning of evolving policy regimes, providing a unified and interpretable approach for sustainable land use planning under non-stationary conditions.