Confronting Challenges in Distribution Modeling for Marten in Western Montana

Abstract

Species distribution models are a popular tool for wildlife management and can be used to identify habitat, guide survey efforts, and predict range shifts under future conditions. Their popularity stems from increased access to occurrence data from public repositories and citizen science platforms. These data present challenges, however, because we often lack information about sampling design and accessible areas are oversampled, potentially leading to biased insights. Mitigating these challenges is rarely straightforward, causing researchers to rely on black-box algorithms that can lead to complex models with poor predictive ability. We incorporated several techniques to address these challenges into distribution models designed to predict habitat for marten (Martes spp.) in western Montana. Marten are a charismatic forest carnivore and an important furbearer species, yet knowledge gaps exist regarding their status and distribution in Montana. We used lasso regression to compare environmental covariates at occurrence records with those at available points. We explored two techniques to address sampling biases: down-weighting occurrence records in regions with high sampling intensity and selecting available locations to reflect the sampling bias in the occurrence records (hereafter, targeted background). Additionally, we used spatially explicit cross-validation to evaluate the model’s predictive performance. The model with down-weighted occurrence records had higher out-of-sample predictive ability than the targeted background approach and predicted values of relative probability of use showed a strong correlation with occurrences. With their growing use for guiding management actions, distribution models require careful consideration of objectives and data limitations to ensure they are effective for wildlife managers.