In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
277 
This second classification process is based on Support Vector 
Machines, which have been applied successfully in the field of 
remote sensing and pattern recognition, e.g. (Vapnik, 1998; 
Fujimura et al., 2008). Whereas we only want to distinguish the 
classes ‘grassland’ and ‘cropland’, it is necessary to split the 
latter into the two classes ‘tilled cropland’ and ‘unfilled 
cropland’, because they appear differently in the data. Hence, 
we have to apply multi-class SVM (Vapnik, 1998) to our 
problem. In the subsequent sections we describe the features 
used in the SVM classification and the actual classification 
process, including the training required for the SVM classifier. 
3.2 Features 
3.2.1 Textural Features: Textural features derived from of 
the co-occurrence matrix can give important hints to separate 
different agricultural classes. We use the Haralick features 
energy, contrast, correlation and homogeneity in our 
classification approach (Haralick et al., 1973). Figure 1 shows a 
scatter plot of texture homogeneity and contrast for the objects 
of a reference dataset. There are relatively clear clusters 
corresponding to grassland and unfilled cropland. However, the 
figure also shows that there is a considerable overlap between 
the cluster for tilled cropland and the others; hence the need for 
additional features that support a clear separation of these 
classes. 
of the line orientations is derived. This histogram is smoothed 
using a Gaussian kernel. All local minima and maxima in the 
histogram are detected and sorted (highest maximum/lowest 
minimum first); if two local maxima are found to be nearly 
coincident (i.e., if they are separated by an orientation 
difference smaller than 45°), the stronger maximum is selected, 
and the smaller one is discarded. The first and second largest 
surviving maxima (Max I and Max 2 ) and the smallest minimum 
{Minf) of the histogram are then used to derive the structural 
features used in the SVM classification. The first structural 
features are Sj = Mini, s 2 = Max h s 3 = Max 2 , the ratio between 
first minimum and first maximum: s 4 = Mini / Max h and the 
ratio between first minimum and second maximum: s 5 = Mini / 
Max 2 . If there is a significant peak in the histogram s 2 will be 
much higher than s 3 , and s 4 will have a smaller value compared 
to a histogram without a significant peak (cf. Figure 3). Another 
structural feature s 5 , also used by Durrieu et al. (2005) is 
derived from the ratio between Max t and Max 2 . s 5 = 1- Max 2 / 
Maxi. If there is a significant first but no significant second 
peak, s 5 will be close to 1, whereas in case there are two peaks 
that are nearly identical, s 5 will be close to 0. The existence of a 
single significant peak in the histogram indicates tilling, 
because our model assumes that there is only one significant 
tilling direction in the GIS object. 
Mean NDVI 
250 
contrast 
« 
14.00 
♦ Tilled Cropland 
12.00 ' * ■ Untilled Cropland 
< Grassland 
in nn 
0,00 
homogeneity 
0,35 0,40 0,45 0,50 0,55 0,60 0,65 0,70 0,75 
Figure 1. Scatter plot of Haralick features contrast and 
homogeneity of objects of a reference dataset. 
3.2.4 Spectral Features: Information about vegetation is 
contained in the infrared band of multispectral images and in 
features derived from it (Ruiz et al., 2004; Hall et al., 2003; 
Itzerott &Kaden, 2007). Similar to the cited works, we use the 
mean and standard deviation of the red, infrared and NDVI as 
spectral features. Figure 2 shows a scatter plot of the mean 
NDVI (scaled from 0 to 255) and the infrared band for the same 
objects as in Figure 1. The spectral features are well-suited for 
separating tilled cropland from unfilled cropland, but the 
clusters for grassland and tilled cropland still overlap. 
3.2.3 Structural Features: A main difference in the appearance 
of cropland and grassland objects in satellite images is caused 
by cultivation, which is conducted more frequently in crop 
fields than in grassland. The agricultural machines normally 
cause parallel straight lines which can be observed in the image. 
The derivation of structural features is limited to the internal 
area of the object, and starts with the extraction of edge pixels 
using the Canny operator (Canny, 1986). These edge pixels are 
transformed into Hough space. In Hough space, parallel lines 
are mapped into points having the same line orientation (/) 
(Figure 3). From the accumulator in Hough space, a histogram 
200 
150 
100 
50 
0 
50 
1 ♦ 
♦ ♦ 
♦ Tilled Cropland 
Grassland 
Mean IR 
100 125 150 175 200 
Figure 2. Scatter Plot of mean NDVI and Infrared of the 
objects of a reference dataset. 
Figure 3. Steps of the structural analysis. 
This approach fails if line structures caused by cultivation are 
not observable (e.g. maize close to harvest, unfilled crop fields), 
if lines in crop fields are not straight respectively parallel to 
each other (e.g. on hillsides), if grassland possesses parallel 
lines (e.g. mowed grassland), and at a specific point in time 
when the crop looks like green grass and structures are not 
visible. The first three problems may be compensated by 
spectral features, though the differentiation between cropland 
and mowed grass may be difficult if the mowed grass (which is