Spatiotemporal Modeling of Spawner-Recruitment Relationships in Eastern Bering Sea Snow Crab (Chionoecetes opilio): From Structured Inference to Deep Learning

Abstract

Snow crab (Chionoecetes opilio) support one of the most economically important shellfish fisheries in the eastern Bering Sea, yet the spatial and temporal processes linking spawning adults to juvenile recruitment remain poorly understood. Spawner-recruit dynamics in this system constitute a biological teleconnection: adult spawning populations regulate the spatial distribution of juvenile recruits through a multi-year sequence of larval dispersal, settlement, and early benthic growth. Identifying this teleconnection requires methods capable of extracting predictive structure from a dataset that is high-dimensional in space but also severely limited in time. This thesis evaluates a gradient of methods for detecting and quantifying the snow crab spawner-recruit teleconnection, spanning classical dimensionality reduction to deep learning. The specific methods evaluated are three statistical approaches with station-level resolution: Empirical Orthogonal Function analysis with Generalized Additive Models (EOF-GAM), Coupled-Factor Generalized Linear Latent Variable Models (CF-GLLVM) with Linked and Spatially Varying Coefficient formulations, and spectral clustering with GAM-based prediction, and a machine learning approach (Vision Transformer, ViT) that is trained on spatially interpolated gridded fields, with Integrated Gradients attribution used to recover biologically interpretable structure from learned predictions. The methods are evaluated using held-out recruitment years spanning 2021-2023, a period of population collapse that provides a rigorous test of out-of-sample generalization. It is not possible to directly compare across methods quantitatively due to differences in spatial data representation. However, the CF-GLLVM achieved the highest station-level Spearman’s correlation (ρ = 0.733) and F1 score (0.782), while the EOF-GAM produced the best hotspot identification (Jaccard = 0.409); spectral clustering performed at or below a climatological baseline across all metrics while the ViT achieved grid-level Spearman’s correlations of 0.83-0.87 and reduced mean absolute error by 33-38% relative to the climatological baseline. Attribution analysis for the ViT revealed a multi-generational thermal signal, consistent with cold pool conditions shaping the thermal environment experienced by larval cohorts at the time of spawning. No single method dominated across all objectives, but together they provide converging evidence of a spatially structured spawner-recruit teleconnection that persists across a period of substantial population change. The framework developed here is broadly transferable to other ecologically and commercially important species where spatially structured leading populations regulate responses in recipient populations across space and time.