Spatially explicit optimization of urban green infrastructure for multiple ecosystem services using deep learning surrogates

Ecosystem services (ESs) are essential considerations in urban planning. Strategically deploying Green Infrastructure (GI) can enhance multiple ESs. However, spatially explicit ES models like InVEST are computationally expensive, limiting their practicality for iterative GI optimization. We present a deep learning (DL)-based optimization framework that replaces InVEST modules with UNet and Attention UNet surrogates. To apply, we trained DL models on selected InVEST ES modules within Maryland and integrated the trained surrogates into a block-scale GI optimization task in Baltimore to maximize urban cooling and habitat quality while minimizing urban nature deficiency and land conversion. UNet demonstrated high accuracy (R² > 0.9 on test data) and reduced optimization time by 95.5 % compared to direct InVEST runs while producing near-identical Pareto-optimal solutions. By enabling rapid, data-driven land-use decisions, this approach provides planners with a powerful tool to integrate ecosystem services into policy and design, facilitating meaningful engagement with non-expert stakeholders.