Uncertainty Characterization in Remotely Sensed Land Cover Information

Uncertainty characterization has become increasingly recognized as an integral component in thematic mapping based on remotely sensed imagery, and descriptors such as percent correctly classified pixels (PCC) and Kappa coefficients of agreement have been devised as thematic accuracy metrics. However, such spatially averaged measures about accuracy neither offer hints about spatial variation in misclassification, nor are useful for quantifying error margins in derivatives, such as the areal extents of different land cover types and the land cover change statistics. Such limitations originate from the deficiency that spatial dependency is not accommodated in the conventional methods for error analysis. Geostatistics provides a good framework for uncertainty characterization in land cover information. Methods for predicting and propagating misclassification will be described on the basis of indicator samples and covariates, such as spectrally derived posteriori probabilities. An experiment using simulated datasets was carried out to quantify the error in land cover change derived from postclassification comparison. It was found that significant biases result from applying joint probability rules assuming temporal independence between misclassifications across time, thus emphasizing the need for the stochastic simulation in error modeling. Further investigations, incorporating indicators and probabilistic data for mapping and propagating misclassification, are anticipated.