Comparing Random K-Fold and Spatial Block Cross-Validation for Predicting Post-Fire Vegetation Recovery

Abstract

Spatial autocorrelation (SAC) can make random cross-validation optimistic for ecological and remote-sensing data, so spatial block cross-validation (SBCV) is widely recommended. How much SBCV changes model assessment depends on how far SAC extends relative to training–test separation. Rather than asking which strategy to use, we varied training–test distances within one dataset and related the resulting difference between SBCV and random five-fold cross-validation (KCV) to the SAC range. We modeled drone-derived vegetation cover (10 m cells) in the second growing season after the 2022 Uljin-Samcheok wildfire, South Korea, with random forest tuned separately per strategy. Vegetation cover was autocorrelated mainly within about 60 m (63% of the variance), a structure that variograms with coarse lag classes can absorb into the nugget. Mainly because almost all KCV test cells had a training cell 10 m away, KCV gave a lower RMSE than SBCV (19.5% vs. 21.9%). The difference shrank as training–test distances increased and vanished after thinning to a 100 m spacing. KCV therefore mainly measures interpolation error near sampled cells, whereas SBCV better approximates prediction at new locations. The need for spatial validation should be judged from autocorrelation at the grid spacing and from training–test distances, not from a nugget-to-sill ratio.