Evaluating the Transferability of Machine-Learning Models for Pre-Emergence Bark Beetle Detection Using Multispectral and Hyperspectral UAV Data

Abstract. UAV-based bark beetle detection often remains site-specific. We assessed transferability using seven DroneNet4Beetles UAV crown-reflectance datasets (multispectral and hyperspectral VNIR) from Sweden, Finland, Italy, and Czechia. We compared Random Forest (RF) and generalized linear model (GLM) using (i) multivariable spectra (multispectral) or multivariable VI sets, and (ii) single-VI GLMs, and evaluated performance along a four-tier ladder (within-plot, cross-plot within stand, cross-stand, cross-country) using Cohen’s kappa over weekly time series. RF and multivariable models generally performed best within-domain, but transferability dropped sharply at the cross-stand level and became most variable across countries. In contrast, the Green-shoulder Curvature Ratio Index (GSCR1, updated formulation in 2026) showed the most consistent transfer between stands and countries, with strong SE↔CZ and CZ↔FI transfer but weak SE↔FI, indicating domain distance—not geographic distance—drives generalization. Transfer involving different hyperspectral sensors was not systematically worse, but multispectral cross-country transfer was highly sensitive, consistent with red-edge band-centre mismatch reducing VI generalizability.