The paper presents an approach to classification and accuracy assessment of ad-hoc categorical maps based on existing spatial datasets with estimates of continuous forest variables; Pixel level class membership probabilities are estimated using a Bayesian network model.
Full text in HTML
Full text in PDF
Spatially explicit data on forest attributes is demanded for various research with landscape perspective. Existing datasets with estimates of continuous forest variables are often used as the basis for producing categorical forest type maps. Normally, this type of maps are used without knowing their accuracy. This paper presents a Bayesian network model for estimating pixel level class membership probabilities of thus derived categorical maps. Class membership probabilities can be used as a post-classification measure of map accuracy and in the process of map classification affecting the assignments of class labels. The method is applied in mapping deciduous dominated forests on the basis of the k-NN Sweden 2005 dataset in a study area in southern Sweden. The results indicate rather low accuracy for deciduous class regardless of the map classification method: 0.48 versus 0.50 in the maps classified without and with the use of the class membership probabilities given equal deciduous area. When probability-based classification is applied, the level of accuracy varies with the assumed map class proportions. Thus, when deciduous class area corresponding to the National Forest Inventory estimate was used, the accuracy of only 0.35 was obtained for the deciduous map class.