Ali Akbar Abkar 
  
  
mun( cost) 
  
0.3322 
0332 \ 
i 
0331556. 03314 
37 
37 
nsk, and 
— 03318 
man( sk) £ f 
03316 a 
l 
/ 
| 1 1 
51 58 65 n 7 
expand 
ue 
  
  
  
  
  
  
  
expand 20 
Figure 5. Forest and non-forest cover 
map obtained by the isotropic 
expand hypothesis generation. 
(a) the graph of the cost as a function of 
the number of expand operations in case 
the land use map of 1977 is used for 
expansion. It has a min-cost of 0.360612 
at expand-parameter-18. 
(b) the best result of the iterative 
isotropic expand operations fori = 18 
(c) deforestation map of Phrao between 
1977 and 1989 using isotropic expand 
operations. 
Figure 6. Forest and non-forest cover 
map obtained by the soil-constrained 
isotropic expand hypothesis 
generation. 
(a) the graph of the cost as a function of 
the number of expand operations. The 
graph has a min-cost of 0.331556 at soil- 
constrained-expand-parameter = 44. 
(b) the result of the non-isotropic 
expand operations for i = 44. 
(c) deforestation map of Phrao between 
1977 and 1989 on the basis of the soil- 
constrained isotropic expand result. 
Figure 7. Forest and non-forest cover map obtained by the non-isotropic expand with the use of the 
seed objects from the land use map of 1977. 
(a) the graph of the cost as a function of the number of conditional expand operations (N cond. Dilation for 
N= 0...900). The graph has a minimum cost of 0.295039 at N = 801. 
(b) the best result of the non-isotropic expand 
(c) deforestation map of Phrao between 1977 and 1989 on the basis of the non-isotropic expand result. 
35 Overall Evaluation of Results 
The resulting minimum costs of error solutions are compared with classical methods of image classification based on 
per-pixel maximum likelihood (ML) classification followed by merging based on majority operators. The results of the 
ML classification are shown in Fig. 8. 
The available reference land use map of 1989 (ground truth) is also used to create the confusion matrices, in order to 
assess the accuracy of the classified results. The reference data is used to check whether the result of the model comes 
close to a man-made interpretation of the non-forest area, notwithstanding the errors made in man-made interpretations 
of the ground truth map—the ground reference data (or ground truth) cannot be regarded as the absolute truth. 
However, for comparison of the model results with the improved ML classification result, a 3x3 Majority filter is 
applied (8 times) to the result of ML classification to smooth the classified result (Table 1). 
  
14 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.