In: Wagner W., Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
> 
> 
571 
shows stacked imagery composed by the NDVI images calculated 
for four consecutive months from April to July. Analysis of the 
stacked NDVI imagery clearly highlights permanent bare areas 
(low NDVI values) as dark features, contrary to the forested or 
vegetated agricultural areas (high NDVI values), which appear 
in brighter shades of blue and yellow. After obtaining the Red 
Edge NDVI images, a 5-band image containing the stacked NDVI 
images for the months of April, May, June, July and September 
were created in ERDAS Imagine. It was finally rescaled to the 
dynamic range of the RapidEye imagery, which is 12 bit. 
Segmentation and Classification: The 5-band stacked NDVI im 
age was segmented in Definiens eCognition, using the spatial data 
of the LPIS as an input thematic layer. The aim was to aggregate 
into single objects, the image pixels with similar temporal be 
haviour in respect to the vegetation cover. The segmentation was 
performed at high detail to preserve features up to 0.1 ha within 
the imagery; as a consequence the land cover features larger than 
the minimum mapping unit, were over-segmented (Figure 3). 
Figure 3: An extracted part from the original stacked NDVI im 
agery and its segmentation by Definiens eCognition. 
The spatial and alphanumeric data from the LPIS plays an inte 
gral role in the segmentation and classification of the RapidEye 
imagery. As a consequence, the resulting land cover segments 
were coherent with the spatial extent and design of the reference 
parcels of the LPIS. In addition, valuable information regarding 
the type of the land use and the farmer restrictions at reference 
parcel level (stored in the LPIS attribute data), was used in the 
subsequent classification and aggregation of the image segments 
into meaningful land cover features at a higher object level. 
The resulting segments were further classified in eCognition, to 
extract various land cover features. Different variables, such as 
Brightness, Mean value of Red, Relative Border to, Border Index 
and Thematic Attribute, have been used. The exhaustive toolbox 
of eCognition, together with the extensive use of abundant Rapid- 
Eye and LPIS data, gave the possibility to define and extract more 
land cover types thus, enrich the initial simple binary classifica 
tion of vegetated and non-vegetated areas. The land cover types 
were further grouped in GAC, Potential non-GAC and Non GAC 
categories, based of the pre-defined rules. 
6 PRELIMINARY RESULTS 
The first results obtained for the test area of KARD (Figures 4.c 
and 4.d) indicate that non-GAC features can be detected with high 
success rate. The overall thematic accuracy of the land cover 
classification is about 81% (See Table 2). The major confusion, 
which was between natural bare areas and urban areas, was not 
considered critical as both classes, eventually, are classified into 
the same (non-GAC) group. For KARD zone 55.1% (10118.8 ha 
out of 18361.3 ha) is in GAC, 8.8% (1614.2 ha) is in potential 
non-GAC and 36.1% (6628.2 ha) is in non-GAC group. 
Some mixed land cover of bare areas and natural vegetation were 
incorrectly validated as agricultural areas because of the vague 
ness associated with ground truth samples. Unfortunately, due 
Urban 
Areas 
Agri 
culture 
Forest 
Bare area 
(natural) 
Water 
Urban areas 
44 
0 
0 
9 
0 
Agriculture 
0 
141 
0 
0 
0 
Forest 
0 
0 
58 
0 
0 
Bare area 
0 
7 
0 
27 
5 
Water 
0 
0 
0 
0 
26 
Other 
0 
37 
0 
10 
0 
Producer ace. 
1.0 
0.76 
1.0 
0.59 
0.84 
User ace. 
0.83 
1.0 
1.0 
0.69 
1.0 
Overall ace. 
0.81 
Table 2: Producer and user accuracies for the clusters extracted 
by object oriented analysis 
to the limited ground truth taken directly in the field, the valida 
tion of the classification was done solely on the base of informa 
tion obtained from the VHR imagery. Even though having suf 
ficient spatial, spectral and radiometric resolution, the IKONOS 
imagery represents only a single snapshot of the ground, a limi 
tation, which cannot always ensure that the information available 
on the VHR image will be sufficient for a proper interpretation 
of the ground truth. A further validation of the results is planned 
with more reliable ground truth data, available from the annual 
field inspection done by the National Administration on selected 
agriculture parcels from the test zones. 
7 CONCURRENT TESTING 
In addition to the object oriented analysis of stacked red edge 
NDVI images, a pixel based method using automated clustering 
of the Self-Organizing Maps (SOMs) (Ta§demir and Milenov, 
2010) has also been utilised. This SOM based analysis has ex 
ploited information in all bands, i.e., each pixel has a 20-band (5 
RapidEye bands for 4 consecutive months from April to July). 
SOMs are unsupervised artificial neural networks that use a self 
organizing learning algorithm inspired from the neural maps on 
the cerebral cortex (Kohonen, 1997). They are successfully used 
in remote sensing applications due to their two main properties: 
i) providing an adaptive vector quantization of the data samples 
to approximate the unknown density distribution of the data; ii) 
distribution of these quantization prototypes on a rigid lattice 
by preserving neighborhood relations in the data space so that 
high-dimensional data spaces can be visualized in lower dimen 
sions (preferably 2D or 3D). Comprehensive knowledge learned 
by SOMs can be used for cluster extraction and knowledge dis 
covery from large data sets using interactive or automated meth 
ods (Ta§demir and Merenyi, 2009, Ta§demir and Milenov, 2010). 
An automated hierarchical clustering of SOMs based on detailed 
local density distribution, proposed in (Ta§demir and Milenov, 
2010), was used for GAC detection and extraction from the 20- 
band stacked RapidEye imagery. A 50 x 50 SOM was trained 
by Matlab SOMtoolbox and a cluster map, focusing on the land 
cover types of permanent bare areas, water, forest and vegetated 
areas, was extracted. Figure 5 shows the resulting cluster map and 
compares it to the map extracted by the object oriented analysis. 
SOM based approach is unable to capture spatial context such as 
inland grass (for example vegetation within forest) whereas ob 
ject oriented approach is unable to correctly capture small fields 
due to its averaging property. Despite these minor details, the re 
sulting cluster maps are quite similar in terms of GAC detection. 
The SOM based clustering is advantageous because it is a faster, 
semi-automated method which requires much less user interac 
tion than the object oriented segmentation.