Generalized additive models with delayed effects and spatial autocorrelation patterns to improve the spatiotemporal prediction of the skipjack (Katsuwonus pelamis) distribution in the Colombian Pacific Ocean

Abstract

Fish populations respond to environmental change with daily, monthly, and annual time delays, depending on each species life cycle. However, these delays are rarely included in Generalized Additive Models (GAMs) for species distribution that uses time-series data. Therefore, the predictions of these models entirely rely on assumptions of immediate fish response to oceanic factors. Spatial autocorrelation is also an issue for GAMs because datasets of fish occurrence usually exhibit this property, and though it has been progressively considered for modeling, it is still frequently ignored. These problems cause low model performance, unstable predictions and more importantly, wrong conclusions for fisheries management. We built and applied Generalized Additive Models with spatial terms and delayed effects of oceanic covariates (SDE-GAMs) to investigate model performance and prediction power for the spatiotemporal distribution of the skipjack (Katsuwonus pelamis), a species of commercial importance across the Exclusive Economic Zone (EEZ) of the Colombian Pacific Ocean. We used satellite-derived Surface Sea Temperature (SST), Sea Level Anomaly (SLA), and Chlorophyll-a (CHLA) as predictors for the Catch Per Unit of Effort (CPUE), considering monthly delayed covariate effects and spatial terms at intra-annual cycles. We evaluated performance improvement of SDE-GAMs compared to that of traditional GAMs (T-GAMs: only immediate covariate effects) and spatial GAMs (S-GAMs: immediate covariate effects plus spatial terms). The model performance of SDE-GAMs was on average 25.4% higher, while its prediction error was on average 43% lower. One, two and three-month delayed SST effects were the primary drivers of CPUE throughout the intra-annual cycle across the EEZ. SDE-GAMs were able to predict both general patterns and smaller details of the spatiotemporal distribution of skipjack, capturing sub-regional differentiation with high importance for management and decision making.