Fig. 1. Bayesian network representation of the joint probability distribution p(X_1j, X_2j, Y_1j, Y_2j). Shaded nodes represent the observed variables, the nodes without shading the hidden variables. The solid-line arrows represent the relationships between the variables. The numbered dashed-line arrows show the sequence of message propagation for inferring the posterior distribution of Y_2j.

Fig. 2. Study area Kronoberg County.

Fig. 3. Bayesian network model for estimating pixels’ class membership probabilities in deciduous forest mapping problem. The variable D_j^y is deciduous volume, S_j^y coniferous volume, A_j^y age, and C_j^y is forest class. D_j^x, S_j^x and A_j^x represent the k-NN estimates of the respective forest attributes. Shaded nodes represent the observed variables, the nodes without shading the hidden variables. The rhomboid shape of the node indicates that C_j^y is a deterministic function of D_j^y and S_j^y.

Fig. 4. Distribution of pixels by deciduous class membership probability. Unclassified – all pixels prior to classification; MAP61, MAP27 and FVC – only pixels labeled as deciduous using the respective classificationg method. (a) Full range of values; (b) left-hand side truncated at 0.10.

Fig. 5. A fragment of the study area (left bottom corner coordinates x: 1367200, y: 6334550 ; coordinate system RT1990). Unclassified – probability map with no class labels assigned; MAP61, MAP27 and FVC – maps produced using the respective classification method. The bar on the left shows the colors used for different levels of deciduous class membership probabilities in the unclassified map.