Empirical assessment of the temporal transferability of landcover-bird abundance models

Abstract

Despite the widespread use of land cover in modeling species abundance, its predictive reliability over time is not well understood. This study examines the temporal transferability of land cover-based models using Breeding Bird Survey data for 42 U.S. bird species and land cover variables from the United States Geological Survey National Land Cover Database (USGS NLCD). Generalized linear models (negative binomial) were trained on 2001 data, incorporating 14 land cover classes across six spatial scales (200– 3000 m). Model selection via AIC involved consideration of over 3,000 models per species. Land Cover Models outperformed null models in both training and test sets for 28 species, with predictive accuracy improving by 3%–30%. Model performance remained stable from 2004 to 2019, demonstrating strong temporal transferability. While species like the American Goldfinch showed consistently low errors, others like the Bald Eagle performed poorly. These findings highlight the importance of evaluating predictive consistency over time in biodiversity monitoring models.