Geography Reference
In-Depth Information
ROC 0.93
Deciduous Broadleaf Forest
ROC 0.8 2
Evergreen Needleleaf Forest
ROC 0.91
Deciduous Needleleaf Forest
ROC 0.91
Mixed Forests
ROC 85
Open Shrubland
0
ROC 85
Closed Shrubland
0
Fig. 3.14
The ROC curve value of different vegetation class rule
In addition, we draw the Receiver Operating Characteristic (ROC) curve of
each forest classification decision rule using the WEKA. The true positive rate
(sensitivity) is plotted in the false positive rate (1-Specificity) function for different
cut-off points in the ROC curve. Each point in the ROC curve represents a sen-
sitivity/specificity pair corresponding to a particular decision threshold. A test with
the perfect discrimination (no overlap in two distributions) is carried out on the
ROC curve that passes through the upper left corner (100 % sensitivity, 100 %
specificity). The closer to the upper left corner the ROC curve is, the higher the
overall accuracy of the test is. The area under ROC curve (AUC) for evergreen
needleleaved forest, deciduous needleleaved forest, deciduous broadleaved forest,
mixed forest, open shrub land and closed shrub land are 0.82, 0.91, 0.93, 0.91, 0.85
and 0.85, respectively (Fig.
3.14
). The biggest value of AUC is assigned to the
evergreen Broadleaved forest, indicating that the result gained by the evergreen
broadleaved forest should be better than other four models.
3.3.3.2 Validation with the Ground Reference Data
It is difficult to carry out the validation of the large-scale map for all land cover
types in all regions due to lack of reference data that can represent the 'true' land
cover. Gong performed the validation of a global land cover map using the ground-
truth sample land cover data from the global flux site (Gong
2009
). In this study,
the accuracy of the input land use data is high and has been validated in 2000. So
we only need to validate the accuracy of the forest type and grassland type. The
