Data and sampling issues associated with accuracy assessment of landcover change maps produced from multi-temporal image classification

Problems associated with adequately populating a temporal confusion matrix that indicates the accuracy of a landcover change map are examined. The nature and availability of verification data such as high resolution imagery for t₁ and t₂ are discussed including inherent difficulties associated with pre-2000 time periods and seasonal mismatches. The scarcity of change pixels for obtaining robust estimates of the true commission and omission error rates are demonstrated numerically. Commission errors can potentially be estimated well since candidate error pixels are identified on the change maps - i.e., any pixel for which change was identified. Accurately estimating rates of omission errors, however, is particularly problematic unless one has ancillary knowledge about where landcover change was most likely to have occurred. The ability to adequately sample omission change errors decreases as overall classification accuracy increases, but is unaffected by the amount of landcover change in a region, and the correspondence between misclassified pixels at t₁ and t₂.