Tropical Forest Cover Typees Differntiation using Satellite
Optical and Radar Data:A Case Study From Jambi, Indonesia
Accuracy analysis of classified maps
Accuracy assessment is a general term for comparing the known reference data (ground truth) and the corresponding results of classification to assess the classification accuracy. One of the most common means of expressing classification accuracy is the preparation of a confusion matrix or a contingency table. In order to assess the accuracy of different classified maps, a confusion matrix was developed for each one of them. A confusion matrix shows a cross tabulation of pixel of test field sample and the classification results. Using the Crossing option of ILWIS program. Every row in the matrix corresponds to a class in the test field sample; every column corresponds to a class in the classification result. Thus the boxes in the diagonal of the contain the number of pixel per class that were correctly classified. The off-diagonal boxes contain the classification error pixels. The overall accuracy is the sum of all correctly classified pixel in a sample divided by total pixel of a sample.
From the confusion matrices the overall accuracy percentage of classified single images mentioned above the following accuracy percentage found in single images of Landsat TM, SOPT, JERS-1 and ERS-1 were 92.12%, 78.26%, 87.67% and 91.5% respectively. However in the fused images of TM with JERS-1, ERS-1 and TM with both JERS-1+ERS-1 were 93.0%, 93.09%, and 92.72% while for SPOT with JERS-1, ERS-1 and SPOT with JERS-1+ERS-1 were 85.01%, 83.84%, and 89.69respectively. For the final fused image of both optical radar images it was found to be 93.68%.
From the four confusion matrices one for each single image, it is clear that the Landsat TM has highest (92.12%)overall accuracy followed by ERS-1 JERS-1 and SPOT respectively. The lowest overall accuracy was found in the SPOT image. This was due to clods presence in the SPOT image. However it should be noticed that the information (number of classes and type of classes) obtained in all the images are not the same. For instance ERS-1 has less number of classes and less details e.g. ERS-1 has one class of forest (where all forest classes have been merged) while JERS-1 has three classes of forest (Logged-over Forest, Secondary Forest and Rubber). The confusion matrices for the fused images of optical and radar images indicate that the overall accuracy for each map and accuracy for each individual class has improved with fusion. Water has highest accuracy followed by settlements, clear-cut areas, rubber, logged-over forest, agriculture, oil palm and secondary forest respectively.
Accuracy analysis for clear-cut areas
Table (1) shows the result of classification accuracy for clear-cut areas in each map. As a single image Landsat TM has highest accuracy (92.14%) for clear-cut, while SPOT has lowest accuracy (83.08%) which can be attributed to cloud presence. Both radar images have reasonable high accuracy for topography and roughness of terrain. However, in fused maps from optical and radar images the accuracy for clear-cut areas has been improved like other classes.
| No
| MAP
| Accuracy % for clear-cut
|
| 1 |
Landsat TM |
92.14 |
| 2 |
SPOT |
83.08 |
| 3 |
JERS-1 |
89.1 |
| 4 |
ERS-1 |
87.89 |
| 5 |
Landsat TM & JERS-1 |
93.44 |
| 6 |
Landsat TM & JERS-1 |
93.03 |
| 7 |
Landsat TM, JERS-1 & ERS-1 |
93.89 |
| 8 |
SPOT&JERS-1 |
85.24 |
| 9 |
SOPT&ERS-1 |
85.71 |
| 10 |
SPOT, JERS-1, ERS-1 |
90.9 |
| 11 |
TM, SPOT, JERS-1, ERS-1 |
93.93 |
The relationship between radar backscatter and forest stand parameters
The relation between radar backscatter and forest parameters, such as height, diameter at breast height (DBH), basal area (BA) and stand density has been partially a challenging issue in this study,. If these relationship shows high positive or negative correlation, researchers and foresters, might then be able to determine those forest stand characteristics from the radar data. To determine the relationship between radar backscatter and the forest stand parameters (i.e., DBH, BA, crown cover and height). 45 sample plots were collected. The influence of the wavelength on the radar backscatter is related to the penetration capability of the radar energy into forest canopy. Longer wavelength (L-band 23.5cm) has higher ability to penetrate forest canopy than C-band. The results of the signatures from JERS-1 image showed good positive correlation r = 0/75 with basal area and r =0.7 with basal area. Both height and crown cover had no statistically significant correlation r =0.21 and r =0.25 with JERS-1 radar backscatter. It is believed that longer wavelength like L-band (JERS-1) should have a very good relationship with forest stand parameters. However, most of the research work that has been done this issue done with plantatyions not natural tropical forest. ERS-1 (C-band5.6cm wavelength) radar backscatter showed low correlation with all stand parameters (e.g.r =0.38,0.17,0.47) except with crown cover r =0.64.
