Tracking the positional uncertainty in 'ground truth'

When working with remotely-sensed data, the difficulties in accurately locating a single pixel on the ground are well documented. A number of phenomena contribute to this positional uncertainty, which is exacerbated by spatial bias within sensor footprints. The uncertainty propagated means that landscapes mapped for verification purposes can rarely be perfectly matched to the satellite images representing those landscapes. This paper evaluates four complementary methods in sequence, each devised to model or correct one type of positional uncertainty in verification datasets. A single test dataset is treated with all four methods in sequence, and subjected to fuzzy classification at each stage so that the effects of each method can be clearly quantified and compared. These methods include a Monte Carlo approach simulating random spatial errors, as well as several more systematic approaches.