Spatial Generalization under Hierarchical Dependence: Evidence from Sediment Modelling
Predictive models of sediment dynamics are often evaluated using validation schemes that overlook spatial and hierarchical dependence, leading to optimistic performance estimates. The effect of validation design under spatial dependence on model generalization is examined using annual sediment data from Mediterranean basins, with repeated observations nested within spatial units. Model structures with different levels of flexibility are compared under a strict spatial block validation scheme. A large share of between-catchment variability is explained by observed covariates; however, predictive performance decreases when models are applied to basins not included in the calibration dataset. Increased flexibility reduces prediction error but does not improve explained variance across catchments. These results indicate that improved accuracy does not imply transferable process representation, highlighting the importance of spatial validation for robust inference across catchments.
Keywords: sediment modelling spatial dependence hierarchical models spatial validation