eDNA metabarcoding and random forest identify super-connector fish species across mangrove–seagrass–coral reef continuums in the South China Sea

Background Tropical coastal seascapes rely on ecological connectivity between mangrove, seagrass, and coral reef habitats to sustain fish biodiversity and fisheries productivity. However, identifying key fish species that facilitate cross-habitat energy flows remains challenging due to limitations of traditional survey methods in turbid, structurally complex environments. Methods We integrated environmental DNA (eDNA) metabarcoding with Random Forest (RF) machine learning to map fish assemblages across a mangrove-seagrass-coral reef continuum in Hainan Island, South China Sea. Results We detected 234 fish species across 78 families - a 4.03-fold increase over conventional gillnet surveys (58 species, paired t-test, t = 11.34, p < 0.001). Using RF classification, we identified 23 "connector species" with high cross-habitat fidelity, of which three super-connectors - Lutjanus russellii , Siganus fuscescens , and Gerres oyena - accounted for 47% of cross-habitat network flow. Simulated removal of these three species reduced overall network connectivity by 68%. The RF model achieved 90.7% accuracy (Kappa = 0.86) in predicting habitat occurrence from eDNA data. Variable importance analysis revealed a salinity threshold at 28 PSU and a distance-to-reef threshold at 2.5 km as primary drivers of community turnover. Our bipartite network analysis showed significant nestedness (NODF = 47.3, p < 0.001), indicating that generalist species disproportionately maintain seascape connectivity. Discussion These findings demonstrate that integrating eDNA with machine learning provides a scalable, high-resolution toolkit for defining functional connectivity and prioritizing species for ecosystem-based fisheries management in the Anthropocene.