Reliable validation is critical for spatial prediction models, yet conventional approaches often neglect a model's ability to interpret geospatial characteristics. Traditional accuracy metrics such as coefficient of determination (R²) and root mean squared error (RMSE) provide global error estimates but fail to detect spatially structured errors, potentially supports models that mislead underlying geographic patterns. This study develops a degree of geocomplexity (DG) metric to examine the ability of spatial model to capture the spatial characteristics of local patterns by comparing the local complexity pattern of the original data with those of the model residuals. We applied this DG metric to tree biomass prediction in eastern Australia, evaluating ten competing spatial models. Findings from this study demonstrate that high prediction accuracy does not guarantee strong geocomplexity interpretability, as models with high R² values do not necessarily achieve high DG scores. Furthermore, we demonstrate that no single model performs optimally everywhere, with the DG-based optimal model map revealing a complex pattern of spatial interpretability. The developed DG metric provides an effective validation strategy for assessing and selecting spatial models based on localized performance within a geographical context.
Liu et al. (Wed,) studied this question.