International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B4, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
For the semi-variogram feature extraction, three major 
parameters had to be set. As proposed in (Balaguer et al, 2010), 
we use six directions to calculate an omnidirectional semi- 
variogram. To obtain significant information about the 
occurring structures, a radius of 30 pixels with one pixel step 
size was chosen based on the RapidEye images (5m GSD). 
There are further parameters for SVM classification. We 
decided to use the one-versus-one-strategy for the multi-class- 
SVM; the classes are cropland, grassland, settlement, industrial 
areas, deciduous and coniferous forest. In this paper we will 
focus only on the classes cropland and grassland. The parameter 
y for the Gaussian Kernel as well as the parameter v to avoid 
the over-fitting are learnt automatically using a cross validation 
with grid search (Hsu et al., 2010). The used training data are 
visualised in Figure 2. 
  
3 
Figure 2: GIS training objects (settlement (green), industrial 
areas (blue), cropland (turquoise), grassland (pink), deciduous 
forest(yellow), coniferous forest (brown)). 
To transfer the pixel-based classification results to GIS-objects 
there are three parameters. The threshold 7, (see equation 4) 
depends on how many incorrect pixels are likely to appear in a 
cropland or grassland object. For instance, a cropland area can 
be surrounded by shrubs or trees, so it is likely that pixels are 
classified as forest. t, can be set easily by an experienced human 
operator. The parameters 74 and f, are given by the definitions in 
the GIS object catalogue. t, depends on the minimum mapping 
areas of other classes than cropland/grassland and is set to 40 
m; £4 is given directly by the GIS object catalogue by the 
minimum mapping areas of the classes others than 
cropland/grassland, e.g. ‘forest’, ‘settlement’ or ‘industry’. It is 
set to 1 ha. 
Basically, the only parameters which have to be set by the 
human operator are A (textural feature) and 1, (to transfer the 
pixel-based results to GIS-objects). In this publication we will 
focus on the setting of the parameter #,, as experience shows 
that this parameter has a high impact to the verification result. 
The parameter t, is tested using a series of variable values of 7, 
starting from 1, 21046 to t, 210096. The results of this analysis 
are summarised regarding the TA a posteriori in Figure 3 and 
regarding the time efficiency in Figure 4. By setting f, to 1096 
not more than 10% of incorrect pixels (equation 1) are allowed 
within a GIS-object. Therefore, the TA a posteriori is really 
high (nearly all errors can be detected) but the rime efficiency is 
low (nearly all GIS-objects have to be reviewed by a human 
operator). If #, is increased, the setting is less strict. For f, = 
100% the TA a posteriori decreased to the same level as the TA 
a priori (no errors could be detected; cropland 95.2%; grassland 
96.5%) but at the same time the time efficiency is 100% (no 
manual effort for the human operator). Our aim is to find a 
setting which is strict enough to find errors in the GIS data set 
but not too strict, so the human operator can save time using the 
approach. 
  
100 
TA [%] 
98 
    
  
       
96 
      
Thematic accuracy a posteriori 
92 S8 cropland  ® grassland 
; t, [96] 
10 "20 "30 40 * 50 60 0.8 M 100 
Figure 3: Dependency of the TA a posteriori (y-axis) from the 
parameter ?, (X-axis); required are 95% (red line). 
       
100 
TE [%] Time efficiency 
80 —- Mgrassland Scropland 
60 
40 
20 
0 ; ; t, mél 
10 20 30 40 50 60 70 80 90° 100 
Figure 4: Dependency of the rime efficiency (y-axis) from the 
parameter f, (x-axis); required are 50% (red line). 
Because the required TA a posteriori of 95% could be achieved 
already before the verification process, the results regarding the 
time efficiency is used to determine the best value for f,. For the 
threshold #, = 60% the time efficiency achieves the required 
50% for both classes (cropland and grassland) for the first time. 
Therefore, the parameter ?, is set to 60% for the detailed 
analysis in section 4.4. 
4.4 Results 
The evaluation results of our approach in form of confusion 
matrices are summarised in Figure 5 to Figure 7. In all cases the 
required thematic accuracy of 95% was already given (above 
95%). However, the thematic accuracy could be increased to 
98% to 99%, whereas the human operator has to review less 
than 50% of GIS-objects. 
  
  
  
  
  
stem 
Reference Accepted 
Accepted 65.5% (2012) 30.4% (935) 
1.2% (36) 2.9% (89) 
  
  
  
  
Figure 5: Evaluation results of set union of the classes cropland 
and grassland (TA a priori = 95.9%, TA a posteriori = 98.8%, 
time efficiency = 66.7%). 
  
  
  
  
  
    
    
  
stem 
Reference Accepted 
Accepted 77.2% (1016) 18.0% (237) 
  
  
  
Figure 6: Evaluation results of the class cropland (TA a priori = 
95.2%, TA a posteriori = 98.4%, time efficiency = 78.8%). 
  
  
  
  
  
  
    
  
  
stem 
Referenc Accepted 
Accepted 56.7% (996) 39.7% (698) 
0.9% (15) 2.7% (47) 
  
  
  
Figure 7: Evaluation results of the class grassland (TA a priori = 
96.5%, TA a posteriori = 99.1%, time efficiency = 57.6%). 
68