When machine learning extrapolates in space: How local data shape spatial transferability

Reliable spatial prediction requires assessing model behaviour under extrapolation, a setting in which standard validation schemes and performance metrics may be misleading due to spatial autocorrelation and limited environmental support. We study how validation design, environmental novelty, and data support jointly shape predictive performance in spatial risk modelling. Using plant health surveillance data on Xylella fastidiosa in France (2015–2023), we analyse a high-dimensional spatial prediction problem combining climatic, soil, and land-use predictors. We propose a spatially informed modelling framework integrating ensemble-based feature selection under environmental block cross-validation, a spatially weighted variant of XGBoost based on predictor-specific smoothing of spatially structured predictors, and explicit diagnostics of prediction support through the Area of Applicability, and compare it to a Bayesian hierarchical spatial logistic regression benchmark. Motivated by the east–west spread of Xylella fastidiosa , we introduce a spatial transfer evaluation scheme that assesses predictive performance in previously unobserved regions under progressive incorporation of local data. This scheme reveals a non-linear transition from extrapolative to interpolative prediction regimes governed by expansion of the applicability domain. Resulting gains in discrimination and calibration dominate over the more limited stabilizing effect of spatial smoothing near the extrapolation boundary. These results emphasize the need to interpret spatial predictions conditionally on their domain of applicability and highlight the role of validation and support diagnostics in spatial statistical modelling.