Deep learning approaches for synthesizing Land Ecological Security pathways in global land-use scenarios

Land-use scenario projections are essential tools for sustainable development planning. However, conventional integrated assessment model (IAM) approaches face challenges including computational constraints and limited scenario diversity across land degradation risk categories. We present a deep-learning methodology for categorizing and synthesizing Land Ecological Security (LES) trajectories using artificial scenario generation based on SDG 15.3.1 degradation indicators. Through examination of 1202 land-use projections from the CMIP6/ScenarioMIP LUH2 database, we establish critical determinants of land-system evolution, particularly forest conversion dynamics and agricultural intensification processes that characterize distinct sustainability pathways. Subsequently, we trained three generative deep learning models (Variational Autoencoder, Conditional GAN, Recurrent Conditional GAN) to synthesize 30,000 artificial land-use trajectories with minimal computational cost across three pathway classifications: High-LES, Mid-LES, and Low-LES categories, reproducing key variable distributions and cross-variable relationships and passing internal land-balance checks for accounting consistency. Our evaluation demonstrates that the Variational Autoencoder framework delivers superior classification performance. This investigation reveals the capacity of deep learning to augment conventional land-system assessment methodologies, enhancing the accessibility and comprehensiveness of land-system scenario analysis through rapid exploration of policy-relevant pathway alternatives while maintaining consistency with traditional IAM frameworks.