Predictive habitat suitability modeling of Nereocystis luetkeana in San Juan County, WA

Abstract

Nereocystis luetkeana (bull kelp) is a foundational component of nearshore ecosystems in the inland waters of Washington State and holds deep cultural significance for Coast Salish peoples. This study presents the first habitat suitability models focused specifically on N. luetkeana in San Juan County, Washington, developed in response to the need for improved spatial tools to support kelp monitoring and restoration planning in this subbasin. Two binomial Generalized Linear Mixed Effects Model (GLMM) spatial models were fit to 888 training points (293 presence, 595 absence) derived from 2024 Samish Indian Nation kelp canopy data, using the sdmTMB package in R. Oceanographic predictor variables were derived from standard (~1km) and high-resolution (~180m) configurations of the Salish Sea Model (SSM and SSMHR respectively). The habitat suitability models (HSMs) constructed in this study were validated through spatial block cross-validation and independent validation using 2025 snorkel survey data (n=345). Effective fetch and slope were the dominant positive predictors of N. luetkeana habitat in both the standard resolution habitat suitability model (HSMSTD) and high-resolution habitat suitability model (HSMHR), and mobile substrate the strongest negative predictor, together reflecting the dependence of N. luetkeana on stable hard substrate in wave-exposed environments. HSMHR resolved four additional predictor associations not clearly defined in HSMSTD and produced a substantially lower spatial random field standard deviation (σO = 9.10 vs 13.97), indicating finer-resolution hydrodynamic predictor variables helped captured spatial structure that standard resolution variables could not resolve. Cross-validation area under the curve (AUC) score was 0.918 for HSMHR and 0.895 for HSMSTD; independent validation AUC was 0.698 for HSMHR and 0.722 for HSMSTD, with documented spatial and temporal mismatch between training and validation datasets that suggest these values underestimate true predictive performance. At a 0.5 probability threshold, HSMSTD predicted 2,241 ha and HSMHR predicted 1,705 ha of suitable habitat across the 0-12m MLLW depth domain. These results demonstrate that the inclusion of higher-resolution hydrodynamic predictor data meaningfully improves habitat suitability predictions, particularly for site-level applications. The resulting suitability maps provide a spatial framework for prioritizing field surveys and identifying candidate restoration sites for Tribal and state resource managers.