A probabilistic approach to environmental change detection with area-class map data

One of the primary methods of studying change in the natural and man-made environment is that of comparison of multi-date maps and images of the earth's surface. Such comparisons are subject to error from a variety of sources including uncertainty in surveyed location, registration of map overlays, classification of land cover, application of the classification system and variation in degree of generalisation. Existing geographical information systems may be criticised for a lack of adequate facilities for evaluating errors arising from automated change detection. This paper presents methods for change detection using polygon area-class maps in which the reliability of the result is assessed using Bayesian multivariate and univariate statistics. The method involves conflation of overlaid vector maps using a maximum likelihood approach to govern decisions on boundary matching, based on a variety of metrics of geometric and semantic similarity, The probabilities of change in the resulting map regions are then determined for each class of change based on training data and associated knowledge of prior probabilities of transitions between particular types of land cover.