X-B4, 2012
OPLAND
over.de,
uality control has
nation is required
f how this can be
lity assessment of
ach combines the
Itispectral as well
textural (Haralick
el-based result is
IS database. The
id two 3-channel
ilts show that by
es time using our
' will present our
and grassland of
c accuracy. The
ct objects in the
IS cropland and
em with the real
nput data into the
images taken in
1e system verifies
lysis approaches
natic comparison
ision whether an
from the system;
system; labelled
re passed on to a
| not have to be
'ractive check by
n operator saves
ient because an
the system is not
time efficiency.
jects is done by
i-automatic one.
ent is essentially
ise, classification
/e to be avoided
ors remaining in
ch is to achieve a
abase after the
and to prove that
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
the approach is suitable for quality assessment, an evaluation is
done in section 4. The paper concludes with a summary and an
outlook in section five.
2. RELATED WORK
The focus of this paper is the verification of GIS-cropland and
grassland objects. The publications dealing with the classes
cropland and grassland using multi-temporal images are limited
to the classification task. Therefore, in this section we will focus
on approaches dealing with the classification of cropland,
grassland and similar classes like vineyards using a multi-
temporal data set with low resolution images. The special focus
will be on features and on the classification method.
Using a multi-temporal data set with low resolution images, it is
common to use only spectral features for the classification
process (Gong et al., 2003; Itzerott and Kaden, 2007; Marçal
und Cunha, 2007; Hall et al., 2008). Itzerott and Kaden (2007)
use for the classification of different agricultural classes norm-
curves of these classes which were created from a prior multi-
temporal analysis based on the Normalised Difference
Vegetation Index (NDVI). These NDVI-norm-curves show that
grassland objects always have an NDVI significantly larger than
zero, whereas cropland can have a very low NDVI depending
on the season. The norm-curves are created using four Landsat
images (GSD: 30m) taken within one year. For the classification
of unknown GIS-objects of a given field boundary cadastre the
mean NDVI of each object is calculated. Then a classification is
carried out using the NDVI norm-curves within a Maximum-
Likelihood or box classification. Using the box classification an
overall accuracy of 65.7% and using the Maximum-Likelihood
classification an overall accuracy of 72.8% could be achieved.
However, the NDVI of different crops can underlie strong
regional and temporal variations. Hence, the adaption of the
NDVI-norm-curves to other regions is a challenge. Training
with a multi-temporal data set within a large area would be
necessary.
Simonneaux et al. (2008) apply a pixel-based approach using a
decision tree algorithm for the classification of different kinds
of crops. For each pixel a NDVI profile over time is calculated.
To create these profiles, eight Landsat satellite images taken
within one year were available. The overall accuracy of this
approach is 83.7%; the kappa-index is 0.78. These good results
could be achieved mainly through the high number of images.
Margal and Cunha (2007) use the NDVI and a field boundary
cadastre for the detection of vineyards in a multi-temporal data
set consisting of nine SPOT 5 images (GSD: 5 m) taken in
2002, 2003 and 2005, and in addition four Chris Proba satellite
images taken in 2006 (GSD: 17 m, 18 bands). Besides the
average NDVI value also the minimum, maximum, standard
deviation and the median NDVI per GIS-object was calculated.
Margal and Cunha (2007) summarise in their article that the
features are useable for the classification but quantitative results
are not presented.
Lucas et al. (2007) proposes a rule based classification based on
the software eCognition (Baatz and Schape, 2000). First,
segments (fields) are determined by a segmentation of each
GIS-object. Next, numerical decision rules based on fuzzy logic
are developed to discriminate vegetation classes. The rules are
primarily based on inferred differences in phenology, structure,
wetness and productivity. The decision rules connect
knowledge about ecology and the information content of single
and multi-temperal remotely sensed data and their derived
products (e.g., vegetation indices). The rule-based classification
65
gives a good representation of the spectral and temporal
characteristics of different agricultural classes but leads to quite
complex rules. These complex rules are difficult to manage and
the transfer to other regions.
De Wit and Clevers (2004) apply a pixel-based Maximum-
Likelihood classification combined with an object-based
decision tree classification. In the pixel- and object-based
classification the NDVI was used as feature. The image data set
used in (De Wit and Clevers, 2004) consist of in total 13
Landsat, two IRS-LISS3 (GSD: 25 m) and two ERS2-SAR
images taken within two years. The overall accuracy of this
approach is high with 90.4%. However, for the object-based
classification first the interactive creation of a field boundary by
a human operator is necessary. Due to the time-consuming
generation of the field boundary cadastre, further improvement
for a practical use of this approach is necessary.
If images of higher resolution are available, additional features
like textural or structural features can be introduced into the
classification process. Textural features describe the distribution
of grey values; structural features describe structures within a
GIS-object such as parallel lines within a local neighbourhood
of a pixel or within a GIS-object. For instance, Miiller et al.
(2010) use spectral, textural and structural features in a
classification based on weighting functions to differentiate
between several kinds of crops. They use high resolution multi-
temporal aerial images (GSD: 17cm). First, the phenological
behaviour of different crops is trained using a training data set
and the determined features. Based on this training, GIS-objects
with an unknown class can be classified by analysing their
phenological behaviour. The results are promising with an
overall correct classification rate of 91.3% but due to the small
size of the test area no final conclusions about the practical
usefulness can be made.
Our method differs from the cited approaches by the used
features, classification method and number of images needed for
the classification/verification. For instance, we use structural
features derived from a semi-variogram. For classification we
use is the state-of-the-art algorithm of Support Vector Machines
(SVM; Vapnik, 1998) which has not been used for the multi-
temporal classification of the agricultural classes cropland and
grassland so far. In addition, to avoid the use of a field
boundary cadastre, we apply a pixel-based classification. Our
approach is flexible regarding to the number of images, and also
can operate with only three images taken in one year.
3. APPROACH
The idea of the approach is to use the fact that the appearance of
cropland changes significantly within a year (cropland can be
covered with vegetation or is not covered with vegetation, it can
contain structures when tilled or not when untilled, ...) whereas
the difference in the appearance of grassland changes only
slightly. As mentioned above, these multi-temporal
characteristics are considered in a pixel-based classification
approach. In order to process the classification three main steps
are necessary. First, features are extracted within a local N, x N;
neighbourhood. Second, these features are classified by a
previously trained supervised learning method. Finally, the
pixel-based results are transferred to the GIS-objects. The object
boundary polygons are given by the GIS data set which has to
be verified.
3.1 Feature Extraction
The feature extraction process takes into account several
different aspects to ensure an optimal classification result. For
