Comparison of machine learning interpolation models for movement trajectories of desert bighorn sheep

Abstract

Global positioning system (GPS) collars have revolutionized wildlife monitoring by enabling fine-scale inference of animal movement, behavior, and habitat use. However, GPS fix failures associated with terrain, canopy, or animal behavior introduce bias into downstream ecological analyses. In a study of 16 translocated desert bighorn sheep ( Ovis canadensis mexicana ) in Sonora, Mexico, we recorded 55,277 scheduled GPS fixes over 18 months, of which 720 (1.3%) were missing not completely at random with the highest rate of missingness during July (268 missing of 3720 scheduled or 7.20%). To interpolate missing locations, we evaluated five predictive models: linear regression (LM), random forest (RF), treed Gaussian processes (TGP), Bayesian additive regression trees (BART), and generalized additive models (GAM). Models were fit using temporally embedded covariates derived from environmental, temporal, and movement-related features. Random forest models achieved the highest accuracy on average (mean Euclidean distance = 54.6 m) across all individuals, outperforming BART (188.3 m), GAM (190.8 m), LM (190.6 m), and TGP (210.5 m). However, RF models also exhibited higher variability in predictive performance. Our results demonstrate that temporally embedded features capture complex behavioral and environmental dependencies, enabling accurate interpolation of GPS fixes. While RF models offer predictive advantages, post hoc averaging with lower-variance models (e.g., BART) may improve uncertainty quantification. These findings provide a demonstration of estimating a phenomenological interpolation model for subsequent animal movement analyses.