Improving land cover change estimates by accounting for classification errors

In monitoring land cover change by overlay of two maps from different dates, the rate of change is frequently overestimated. This is due to three sources of uncertainty: (1) semantic differences in class definitions between two maps, (2) positional errors and (3) classification errors. In this study, four methods are proposed that use the Bayes theorem to update prior estimates of land cover change with information on the probabilities with which land cover classes are mistaken for each other. The methods were illustrated for two case studies. In the first case study, the real change was 1.4%, and by overlay of the two maps, 7.4% was predicted. The estimates by the four methods were 6.3%, 15.3%, 6.7% and 1.6%. In the second study, these percentages were 48% and 36%, and with our four methods: 39.2%, 54.1%, 50.8% and 53.0%. Two of the methods account for the correlation in classification accuracy between maps of two dates. Where this correlation was high (study area 1), the methods that accounted for correlation yielded change estimates closer to the real change than the methods that did not account for this correlatio