AI-enhanced urban forecasting: ConvLSTM networks for multi-scenario land cover prediction in metropolitan regions

Abstract Long-term urban growth prediction requires capturing complex spatiotemporal dependencies that emerge over multi-decadal timescales. This study develops a Convolutional Long Short-Term Memory (ConvLSTM) framework applied to a uniquely dense 39-year annual land cover dataset (1985-2023) across Colorado’s Metropolitan Planning Organization areas to address limitations in long-term spatiotemporal urban growth prediction. The framework integrates stakeholder-driven scenario planning to evaluate five alternative development trajectories ranging from business-as-usual to high-growth futures. Analysis reveals a counterintuitive “Temporal Depth Paradox” where prediction accuracy (F1 score) increases substantially from 0.27 for 1-year forecasts to 0.87 for 20-year forecasts, challenging fundamental assumptions in urban forecasting. The model effectively distinguishes between different development processes, showing higher predictive accuracy for greenfield expansion versus complex urban redevelopment. By embedding ecological constraints and evaluating alternative growth scenarios, the framework demonstrates how AI can support sustainable development goals while maintaining human oversight in planning processes. Quantitative scenario analysis reveals that compact development patterns (Adaptive Innovation scenario) produces 20.5% infrastructure efficiency improvements and preserve approximately 1,000 additional hectares of natural areas by 2050 compared to unregulated growth scenarios. The findings provide planners with evidence-based tools to evaluate how different policy interventions might reshape urban morphology, enhance infrastructure efficiency, and preserve natural resources in rapidly growing regions.