Modeling fish habitat: model tuning, fit metrics, and applications

Knowledge of the habitat associations and spatial–temporal distributions of wild animals is essential for successful ecosystem management, and effective analytical approaches are key to develop accurate models of these relationships. We explore the influence of several modeling techniques, tuning parameters, and assignment thresholds on a variety of model fit metrics to characterize habitat associations and make spatial–temporal predictions of species distribution based on a nine-year acoustic telemetry fish tracking dataset from a freshwater system. Unweighted generalized linear mixed models (GLMM) and random forests (RF) had the highest prediction accuracy of fish occupancy (> 84%) and precision (positive predictive value accuracy), but because the data were imbalanced (> 70% absences), predictions had low sensitivity (accuracy of true presences, < 45%), and therefore, low accuracy balance. Model weighting to prioritize presences and lowered presence probability thresholds both produced more balanced models, but RF exhibited low sensitivity to alterations in probability thresholds. Model weighting presents a straightforward approach to balance class accuracy in imbalanced datasets, which are common in species distribution samples. However, there is a wide range of weighting options and an important trade-off between model sensitivity and precision, either of which may be favoured depending on the research question or management application.