Validation Criterion Functions for Geospatial Classification: A Case Study of Election Data

Abstract

Classification on geospatial data is different from classical classification in that spatial context must be taken into account. In particular, the validation criterion functions should incorporate both classification accuracy and spatial accuracy. However, direct combination of the two accuracies is cumbersome, due to their different subjects and scales. To circumvent this difficulty, we develop two new criterion functions that indirectly incorporate spatial accuracy into classification accuracy-based functions. Next, we formally introduce a set of ideal properties that an appropriate criterion function should satisfy, giving a more meaningful interpretation to the relative significance coefficient in the weighted scheme. Finally, we compare the proposed new criterion functions with existing ones on a large data set for 1980 US presidential election.