Integrating disparate survey data in species distribution models demonstrate the need for robust model evaluation

Abstract

Marine spatial planning and conservation initiatives benefit from an understanding of species distributions across larger geographic areas than are often sampled by any one survey. Here, we test whether the integration of disparate survey data can improve habitat predictions across a region not well sampled by a single survey using Dungeness crab ( Metacarcinus magister) from British Columbia as a case study. We assemble data from dive, trawl, and baited-trap surveys to generate six candidate generalized linear mixed-effect models with spatial random fields. To compare single-survey and integrated models, we evaluate predictive performance with spatially buffered leave-one-out cross-validation and independently with two novel approaches using fisheries catch data. We find improved predictive performance and reduced uncertainty when integrating data from surveys that suffer from small sample size, low detectability, or limited spatial coverage. We demonstrate the importance of robust model evaluation when integrating data and predicting to unsampled locations. In addition, we highlight the need for careful consideration of sampling biases and model assumptions when integrating data to reduce the risk of prediction errors.