Disentangle classifier error from spatial-reference-mask error

Determine how much of the pixel-level disagreement between the XGBoost wheat classification product and the government spatial reference mask reflects classification error and how much reflects error in the positively biased spatial reference mask.

Background

The operational XGBoost classifier’s predicted wheat area differed from the official tabular target by only +2.99%, whereas the government spatial reference mask exceeded the same target by +25.11%. Precision and recall for the deployment-scale classifier were nevertheless computed relative to that spatial mask.

Because the spatial mask is itself demonstrably biased and no independent probability sample of field-verified points was available at district scale, the paper states that the source of the pixel-level disagreement cannot be determined. Resolving this issue would require an independent, field-verified reference design capable of separately assessing the classifier and the administrative spatial product.