How to Evaluate ML Models in Geospatial Settings?
Published:
Standard K-fold Cross-validation (KFCV) randomly divides a training set into K non-overlapping folds and iteratively holds out one fold at a time, training a model on the remainder (i.e., training folds) and measuring error on the held-out fold (i.e., validation fold). The average of these model errors across folds is the estimate of generalization performance for an unseen test set. KFCV provides unbiased performance estimates when applied to independent, identically distributed (iid) data. But does it also work well on geospatial data?