385 
ENHANCING THE AUTOMATIC VERIFICATION OF CROPLAND 
IN HIGH-RESOLUTION SATELLITE IMAGERY 
P. Helmholz a,b *, F. Rottensteiner b , C. Fraser b 
institute of Photogrammetry and Geoinformation, Leibnitz University Hannover, Nienburger Straße 1, 
D-30167 Hannover, Germany - helmholz@ipi.uni-hannover.de 
b Cooperative Research Centre for Spatial Information, Dept, of Geomatics, University of Melbourne, 
723 Swanston Street, Carlton VIC 3053, Australia - (petrah, franzr, c.fraser)@unimelb.edu.au 
Commission IV, WG IV/3 
KEY WORDS: Segmentation, Grouping, Image Understanding, Feature Extraction 
ABSTRACT: 
Segmentation is one of the first steps in the field of image analyzing and image understanding. It is the basis for the interpretation of 
images, or it supports other techniques which is the motivation for the in this paper introduced segmentation algorithm. In this paper, 
a segmentation algorithm and its application to enhance an approach for the automatic verification of tilled cropland objects in a GIS. 
For this application, cropland objects in a GIS that may contain more than one field should be segmented into individual 
management units. The algorithm starts with a Watershed segmentation that results in a strong over-segmentation of the image. A 
region adjacency graph is generated, and neighbouring segments are merged based on similarity of grey levels, noise levels, and the 
significance of the boundary between the segments. After segmentation, the verification algorithm can be applied separately to the 
individual segments, and finally, these verification results have to be combined, taking into consideration the specifications of the 
GIS. Several examples show how the segmentation process can help to improve the verification of the cropland objects in the GIS 
from IKONOS images covering a test area in Germany, but also the limitations of the segmentation algorithm. 
1. INTRODUCTION 
The goal of the project WiPKA-QS (Wissensbasierter 
Photogrammetrisch-Kartographischer Arbeitsplatz 
Qualtitätssicherung; Knowledge-based photogrammetric- 
cartographic workstation - Quality control) is the automatic 
verification of ATKIS (Amtlich topographisch-kartographisches 
Informatisonssystem - Authoritative Topographic-Cartographic 
Information System) or other comparable Geographic 
Information Systems (GIS) by comparing the GIS with high 
resolution satellite imagery (Busch et al, 2004). The main 
components of ATKIS are the object-based digital landscape 
models with a geometrical accuracy of up to ±3 m. 
In order to achieve this goal, a series of algorithms is being 
developed, each aiming at the verification of a specific object 
class defined in the GIS. One of the object classes of interest in 
this context is the class cropland. In Helmholz et al. (2007) an 
algorithm was introduced to verify tilled cropland objects using 
characteristic structural features (parallel straight lines) that are 
generated by agricultural machines during the cultivation. 
These features are observable in satellite images having a 
resolution of 1 m or better. The approach works in three steps. 
For any particular object of class cropland in the GIS, edges are 
detected in the image area enclosed by the boundary polygon of 
the object. Then, the edge image is transformed into the Hough 
space. Finally, after the determination of points of interest (POI) 
in Hough space, a histogram is calculated. The histogram 
represents the number of occurrences of POIs in Hough Space 
(equivalent to lines of interest in the image space) depending on 
the angle. If a significant orientation is determinate by a 
statistical analysis of the histogram, the cropland object is 
accepted. Otherwise, the cropland object is rejected by the 
system and highlighted as an object that has possibly changed 
in its land cover or use. The highlighted objects are to be 
checked by a human operator. 
One specification of ATKIS is that inside a cropland object the 
existence of more than one land cover class is tolerated if a size 
threshold is not exceeded. Several fields of the same land cover 
type (different management units) are also permitted inside a 
cropland object. The existence of more than one management 
unit negatively affects the edge detection process due to strong 
differences of the image properties in the different management 
units. As a consequence, the approach for the detection of 
parallel lines often fails on GIS objects that contain more than 
one management unit (Figure 1). 
In order to overcome these problems it is necessary to split the 
cropland object in the GIS into segments having homogeneous 
grey level properties before carrying out the actual verification 
process. The individual segments are likely to correspond to 
different management units provided they have a certain 
minimum size in object space. After the segmentation, the 
verification algorithm can be applied independently to the 
individual management units, which makes it more robust with 
respect to the automatic tuning of parameters. Afterwards, the 
verification results of all segments are merged and a final 
assessment of the GIS object is done; in addition, due to the fact 
that an individual classification of the homogeneous segments 
has been carried out, the areas of possible change (i.e., 
segments found no longer to be cropland objects) can also be 
highlighted. 
Corresponding author.