Variation in seagrass habitat use by fishery-important nekton across the Gulf of Mexico revealed by deep transfer learning with trophic group priors
Understanding how species use habitats across environmental gradients is central to guiding fisheries management and habitat restoration, yet inference is often limited by heterogeneous data and inconsistent observations. In coastal ecosystems, variation in relationships between nekton and seagrass habitats remains unresolved, in part because habitat structure, environmental context, and sampling methods are rarely integrated in predictive models. Here, we combine multi-gear monitoring data from across the Gulf of Mexico with a transfer-learning framework that incorporates trophic priors to quantify how nekton respond to seagrass structure under varying environmental conditions. We show that environmental gradients--including temperature, salinity, and water clarity--define broad-scale distributions, while seagrass structure refines habitat use at local scales. Apparent inconsistencies in seagrass-nekton relationships are largely attributable to environmental context and differences in observational processes. By integrating observations collected using different sampling methods, our approach reveals consistent species-environment relationships across sites and improves predictive performance, particularly for data-limited species, by using trophic priors. We further show that species differ in their responses to environmental gradients, with some exhibiting consistent patterns across sites and others showing strong context dependence. These results demonstrate that combining heterogeneous datasets can strengthen ecological inference and provide a pathway for scalable, data-driven conservation and restoration in rapidly changing coastal systems.