Comparing Classical, Hidden-markov and Deep-learning Step-selection Functions for Asian Elephant Habitat Preference: The Value of Feature Engineering, Dynamic Diel/Seasonal Variables and Water-quality Covariates

This methodological case study compares differences and nuances among three step-selection function (SSF) families: local-Gibbs HMM-RSF (steady-state utilisation surface), hidden Markov SSF (HMM-SSF, state-conditional β plus simulated visit density) and deep convolutional SSF (deepSSF, per-pixel selection-probability surface). Each is read against ecological prior knowledge to motivate usage recommendations rather than rank methods. We use 16 502 two-hour GPS fixes from two solitary adult male Asian elephants (Elephas maximus) in Kedah, Peninsular Malaysia (2013–2016); at n=2 we report which method recovers which signal under identical inputs, not population-level inference. A unified 91-raster product covers topography, canopy, hydrology, water quality, human pressure and lacunarity vegetation classes at ~10m. Six deepSSF variants (F1–F6) are evaluated not by likelihood but by whether each captures elephant priors: diel canopy use, water-driven routines, human-activity avoidance, July–October fruiting + plantation-activity window. Across 9 strata and 8 (quarter × diel) cells, deepSSF variants agree in forest interiors (ρ ≈ 0.80) and diverge in human-modified strata; F2 and F6 reproduce the diel signal in directions consistent with elephant ecology, while F3 collapses it. The local-Gibbs agrees with deepSSF most in forest_interior, the lowest-σ_seasonal stratum; HMM-SSF's TPM landform predictors expose a movement vs. foraging axis complementing deepSSF's selection surface. On the 10km DynQual water-quality stress test local-Gibbs returns Pathogen β = +3.74 (CI excluding zero on the wrong side), HMM-SSF returns label-switching values and deepSSF averages a monotone-decreasing Pathogen marginal-response across 25 multi-seed checkpoints (mean Δ at p95 = −0.018, envelope including zero): at this coarse resolution the linear-predictor method is confidently wrong, the convolutional method correct on average but uncertain. We recommend matching method to research goal (steady-state utilisation, latent behavioural state, or high-dimensional input flexibility), adopting from each method what is methodologically sound for the question at hand.