Approximate is Enough: Distance-Based Validation for Geospatial Classification

In geospatial classification, the validation criterion functions should incorporate both error rate based classification accuracy and distance based spatial accuracy. However, due to the difference in subject and scale between the two accuracies, it is not trivial to combine them in a reasonable way. To circumvent this difficulty, we develop approximate functions for spatial accuracy that preserve distance ranking instead of distance values. The resultant criterion functions not only take less computation cost, but also get more commensurate with classification accuracy. Finally, the approximation power of the proposed criterion functions are validated on real-world datasets.