Constraint-based graph optimization for marine protected area networks: A graph neural network-informed approach

Designing cost-effective Marine Protected Area (MPA) networks that balance biodiversity conservation, ecological connectivity, and socioeconomic feasibility presents significant computational challenges, particularly at basin scales involving thousands of candidate sites. Traditional systematic conservation planning approaches either rely on simplified spatial prioritization metrics or face computational intractability when incorporating explicit connectivity constraints. We present a novel computational framework integrating Graph Convolutional Networks (GCN) with constraint-based greedy optimization to address this multi-objective problem. The GCN learns compact 16-dimensional site embeddings from a spatial connectivity graph representing 10,741 candidate sites (30 km × 30 km resolution), compressing biodiversity value, cost-efficiency, and network topology into an interpretable representation for exploratory analysis of site trade-offs. A two-phase greedy algorithm, operating directly on the underlying connectivity matrix, first satisfies mandatory ecological constraints (threatened species protection, bathymetric representation, regional equity), then maximizes marginal efficiency defined as biodiversity and connectivity gains per unit cost. We demonstrate the framework using 84,721 species observations covering 120 marine taxa across the Mediterranean Sea, combined with bathymetric data and spatially explicit opportunity costs. For a 15% coverage target (1611 sites, 1449,900 km²), our approach achieves network connectivity 4.4-fold higher than random selection and 1.6-fold higher than biodiversity hotspot strategies, at total cost of €1.69 billion—42% below random selection and 58% below hotspot approaches—while guaranteeing protection of all nine threatened species (four Critically Endangered, three Endangered, two Vulnerable) alongside near-complete overall coverage (99.2% of 120 species), a guarantee not met by cost-minimizing alternatives. All ecological constraints are satisfied while maintaining computational efficiency (8–15min runtime). Network configuration remains robust across coverage-target scenarios (10%, 15%, 30%): species coverage remains stable at 99.2%, and regional distribution varies by <2 percentage points across scenarios. The modular framework is transferable to other marine regions through substitution of region-specific biodiversity, environmental, and socioeconomic data, supporting systematic conservation planning where computational efficiency is essential for stakeholder engagement.