Thematic accuracy assessment of geographic object-based image classification

Geographic object-based image analysis is an image-processing method where groups of spatially adjacent pixels are classified as if they were behaving as a whole unit. This approach raises concerns about the way subsequent validation studies must be conducted. Indeed, classical point-based sampling strategies based on the spatial distribution of sample points (using systematic, probabilistic or stratified probabilistic sampling) do not rely on the same concept of objects and may prove to be less appropriate than the methods explicitly built on the concept of objects used for the classification step.

In this study, an original object-based sampling strategy is compared with other approaches used in the literature for the thematic accuracy assessment of object-based classifications. The new sampling scheme and sample analysis are founded on a sound theoretical framework based on few working hypotheses. The performance of the sampling strategies is quantified using simulated object-based classifications results of a Quickbird imagery. The bias and the variance of the overall accuracy estimates were used as indicators of the method's benefits.

The main advantage of the object-based predictor of the overall accuracy is its performance: for a given confidence interval, it requires fewer sampling units than the other methods. In many cases, this can help to noticeably reduce the sampling effort. Beyond the efficiency, more conceptual differences between point-based and object-based samplings are discussed. First, geolocation errors do not influence the object-based thematic accuracy as they do for point-based accuracy. These errors need to be addressed independently to provide the geolocation precision. Second, the response design is more complex in object-based accuracy assessment. This is interesting for complex classes but might be an issue in case of large segmentation errors. Finally, there is a larger likelihood to reach the minimum sample size for each class with an object-based sampling than in a point-based sampling.

Further work is necessary to reach the same suitability than point-based sampling for pixel-based classification, but this pioneer study shows that object-based sampling could be implemented within a statistically sound framework.