Confidence in coincidence

Abstract

We present a simple simulation scheme to derive confidence intervals for measures computed based on a coincidence matrix. Our approach is based on the conditional distributions between two categorical maps, and is a direct interpretation of how much information one map (usually the classified image) carries about the other map (usually the reference image). The simulation algorithm creates a realization of the map created by visiting each pixel and drawing a random sample from the conditional distribution of reference categories (conditioned on the category of the pixel of the classified image). Confidence intervals can be derived by repeating the simulation many times and computing the coincidence measure(s). Handling the coincidence matrix as a set of conditional distributions also allows the comparison of maps with different numbers of categories. This approach is an extension of the traditional methodology widely used in accuracy assessment of data derived from remotely sensed images. We illustrate the usage and interpretation of the approach on artificial and Canadian land cover mapping data.

Acknowledgement

This research was funded in part by the GEOIDE Network of Centres of Excellence and the Natural Science and Engineering Research Council of Canada.