In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
the combined class cropland/grassland. The results reported in 
(Rengers & Prinz, 2009) and (Busch et al., 2004) show that a 
purely textural analysis is not sufficient for separating cropland 
and grassland. Spectral and / or structural information is 
required for that purpose. 
Haralick et al. (1973) used textural features derived from the 
grey level co-occurence matrix such as energy, contrast, 
correlation and entropy are used along with the mean and 
standard deviation of the gray values of all four available 
channels to classify coastline, forest, grassland, urban areas and 
irrigated and non-irrigated cropland using a linear discriminant 
function method. By combining a textural analysis with the 
spectral features the classification accuracy could be improved 
over a purely radiometric analysis. More recently, Itzerott and 
Kaden (2007) tried to distinguish various types of farmland 
using solely the Normalised Difference Vegetation Index 
(NDVI) that is computed from the near infrared and the red 
bands of a multispectral image. Analysing typical crops and 
grassland in Germany, they could show that grassland possesses 
an NDVI that is significantly larger than zero in all seasons, 
whereas untilled cropland has a very low NDVI except for a 
short period. However, they observed strong regional and 
temporal variations of the NDVI, so that statistical parameters 
describing the NDVI of the different agricultural classes in one 
region are hard to transfer to other regions. Training with a 
multitemporal dataset within a large area would be necessary. 
Hall et al. (2003) use the NDVI to separate vines and bare soil 
in aerial images with a spatial resolution of 0.25 m. Afterwards, 
the orientation of the rows is calculated using a priori 
knowledge about the distances between the rows and between 
the individual plants within a row. However, such a priori 
knowledge is usually not available for cropland objects. 
Structural features have been used frequently to distinguish 
agricultural object classes such as vineyards, orchards, or 
plantations. The structural characteristics exploited for the 
extraction of these objects, namely straight parallel lines, also 
occur in cropland, where they are caused by tilling. However, 
some assumptions usually made in the extraction of vineyards 
or orchards cannot usually be made for cropland. For instance, 
Chanussot et al. (2005) estimate the orientation of vineyard 
rows automatically from aerial images by using the Fourier 
spectrum of an image and its Radon transform. Wassenaar et al. 
(2002) detect orchards and different kinds of vineyards in aerial 
images using a Fast Fourier Transformation, using specific 
knowledge about the distances between vine rows to reduce the 
search space in the frequency domain. Delenne et al. (2008) use 
a frequency analysis based on Gabor filters to estimate the row 
width and orientation and to detect the boundaries of vineyards. 
All these methods assume the rows of vines to be approximately 
equally spaced or even utilize knowledge about the actual 
spacing of these rows. Both assumptions cannot be made for 
cropland. In cropland the distance between furrows can vary 
from one field to the next depending on the type of crop planted 
in the field, on the kind of machine used for tilling, and on the 
visibility of the structures in the image. 
Trias-Sanz (2006) uses only structural features to discriminate 
objects with similar radiometric and textural properties, namely 
cropland, forest, orchards, and vineyards. These object classes 
can be distinguished only by orientation characteristics. A small 
window is extracted randomly inside an object to be classified, 
and this window (called texton) is used to compute a variogram 
of the image. A histogram of direction angles is derived from 
the Radon transform of the variogram. The maximum of this 
histogram corresponds to the primary direction of edges in the 
image, and it is used in the classification process. The approach 
can be used to discriminate a large number of object classes by 
properly choosing the texton, but can give wrong results if the 
texton size is selected inappropriately. Another disadvantage of 
this approach is that the cultivation structures and field crop 
have to be homogeneous in appearance. Therefore, LeBris and 
Boldo (2007) use a segmentation to extract homogenous 
regions before applying the algorithm of Trias-Sanz (2006). 
A differentiation between agricultural classes such as grassland 
and cropland only on the basis of spectral, structural or textural 
features in monotemporal imagery seems to be impossible. An 
approach which combines these features is introduced by Ruiz 
et al., (2004, 2007) and Recio et al. (2006). Besides spectral 
(mean and deviation of the red, infrared and NDVI channel) and 
textural features determined from the grey level co-occurrence 
matrix (Ruiz et al., 2004), structural features determined from a 
semi-variogram, Hough- and Fourier transformation (Ruiz et 
al., 2007) are used to detect olive trees, citrus orchards, forests 
and shrubs using images of 0.50 m spatial resolution. The final 
decision is based on a decision tree (Recio et al., 2006). In 
addition to the features described so far, information about the 
shape of the object can be use for the classification process. 
Such information can be derived e.g. from a given GIS. 
Hermosilla et al. (2010) extend the approach of Ruiz et al. 
(2007) by using object shape as an additional feature to 
distinguish the classes building, forest, greenhouse, shrub lands, 
arable land and vineyard. Whereas this could improve the 
classification accuracy, it resulted in an increase of the number 
of undetected errors in a GIS to be verified by that approach. 
Our method differs from the cited approaches by the way the 
textural analysis is carried out and by the definition of the 
structural, spectral and textural features. Furthermore, we use a 
different method for classification. The fact that our approach is 
embedded in a system for the verification of GIS objects has 
some implications for the strategy used for classification. The 
parameters of the method have to be tuned according to the 
quality requirements of the GIS: an undetected false 
classification in the GIS is penalized higher than a correct 
classification erroneously highlighted as false. 
3. APPROACH 
3.1 Overview 
The goal of our approach is the separation and verification of 
cropland and grassland GIS objects using 1 m orthorectified and 
pan-sharpened multispectral IKONOS images. In this paper we 
assume that each ATKIS GIS object corresponds to exactly one 
class. The verification process is carried out separately for each 
GIS object. The object’s boundary polygon given by the GIS is 
used to limit the analysis to areas inside the object. In a first 
step we use a supervised classification technique that analyses 
image texture with the help of Markov Random Fields (Muller, 
2007, Busch et al., 2004) to distinguish the combined class 
‘agriculture’, which comprises both cropland and grassland 
objects, from other classes such as ‘settlement’, ‘industry’ or 
‘forest’. If a cropland or grassland object is classified as 
belonging to any other class than ‘agriculture’, it is considered 
to correspond to an error in the GIS. As the algorithm of Busch 
et al. (2004) cannot differentiate between grassland and 
cropland objects, all the other objects (i.e., those passing the 
first classification stage) are passed on to a second classification 
process designed to discriminate grassland and cropland. The 
second classification and the following verification process is 
the main focus of this paper.