maize.” Misclassification was found in the built-up area and
tree area. Most of these misclassified image objects were not
adjacent to each other. Therefore, the entire image objects in
the class “combined maize” were merged together and small
misclassified image objects were eliminated by a query: all
those image objects from the combined maize class that had an
area smaller than 10,000 pixels ought to be unclassified. Due to
the query, most of the small, misclassified image objects were
successfully eliminated from the combined maize class. How-
ever, there were still some areas within other crops that have
been classified as maize (Figure 9a). This misclassification
could not be removed at level 2 image objects. Misclassified
regions were eliminated from the maize class by classifying the
level 1 image objects (field boundaries) with the query: “clas-
sify image objects at level 1 as maize, which have 50 percent or
more area covered by level 2 class ‘combined maize’, other-
wise image object should remain unclassified” (Figure 9b).
Results and Discussion
In the study area, 874 ha were classified as maize from the
total 8,229 ha, which is 10.6 percent of maize of the whole
study area. According to government statistics 11 percent
maize was found in the whole Luechow Dannenberg district.
This study validates the statistical information provided by
the German government. Besides the boom of renewable
energy sources occurred due to political support, 10.6 percent
maize shows that it has not been overly produced. Therefore,
Lucie is effectively helping in protection of endangered birds
species by conserving their habitats. The spatial distribution
of maize has been shown in Figure 10.
To measure the accuracy of classification, an error matrix,
also known as confusion matrix, was created (Story and Con-
galton, 1986; Congalton, 1991). Table 2 shows the error matrix
for grid 738-774.
T
able
2. E
rror
M
atrix
of
G
rid
738
to
774
Maize
Others
Total
User’s Accuracy
Maize
841196.79 100606.69 941803.48 89.32%
Others
153584.80 3201166.80 3354751.60 95.42%
Total
994781.59 3301773.49 4296555.08
Producer’s
Accuracy 84.56% 96.95%
Total = 94.08%
The overall accuracy is the simplest estimator that can be
extracted from error matrix. Overall accuracy of this study was
94.08 percent, which is quite high. It shows kind of biasness
due to the fact that only two classes were used. For detailed
analysis of accuracy of individual classes, user’s and pro-
ducer’s accuracies were calculated. User’s accuracy of maize,
also known as error of commission, was 89.32 percent. About
11 percent of other agriculture land was included in the maize.
Producer’s accuracy of maize, also called error of omission,
was 84.56 percent indicating just over 15 percent of maize
was omitted. Higher user’s accuracy than producer accuracy
demonstrates that the rule set for maize was somewhat strict;
that led to some maize image objects being classified as others
class. The Kappa statistic was calculated to measure the pos-
sibility of correctly classified objects (agreement) by chance
(Cohen, 1960). Its values range from 1 (perfect agreement/accu-
rate) to 0 (agreement by chance/inaccurate). Kappa value was
0.83 that shows the robustness of the methodology adopted.
In this study,
GEOBIA
rules implemented on single date
VHR
imagery (35 cm) clearly resulted in a high user’s and produc-
er’s accuracy. This exhibits that remotely sensed
VHR
airborne
imagery was extremely efficient in mapping the bioenergy
crop in the Natura 2000 area of Lucie. Thirty-five centimeters
(35 cm) of the
GSD
yielded high accuracy in mapping maize,
although the imagery only had information in NIR, Red,
and Green channels. In this research, maize showed diverse
spectral responses as a result of different sowing times and
phenology. Due to vast variation within the class of maize, it
presented spectral signatures similar to some other agriculture
crops, i.e., grassland and rye. The rules developed for split-
ting maize into three subclasses helped in mapping maize
precisely. The object-based classification rules provided the
opportunity to integrate expert knowledge by incorporating
features based on spectral values, i.e., different vegetation
indices from simple ratios to
NDVI
, standard deviations, and
knowledge of crop phenology.
NDVI
enabled for compensating
for variation caused by different illumination conditions and
viewing aspects, and standard deviation allowed to limit or
mask the range of spectral values (Jensen, 2005). Simple ratios
and standard deviations were introduced into classification
because
NDVI
alone could not extract maize accurately. This
is due the fact that
NDVI
is negatively affected by soil surface
reflection within a land cover class (Heute and Jackson, 1988).
In the present study, maize, rye, and grasslands were the
major crops influenced by exposure of soil underneath. That
is why very specific thresholds were devised so that other
crops influence on the maize class is minimalistic, i.e., keep
the error of commission low. The execution time of developed
rules was very fast and helped to improve the accuracy. Stolz
et al
. (2005) implemented
ENPOC
rule based classifier to clas-
sify maize but yielded very low maize accuracy due to similar
spectral signatures of maize and grassland. The present study
helped to overcome this problem by dividing maize into three
subclasses. Hernando
et al.,
(2012b) also mapped Natura
2000 habitat by applying rules (based on threshold values) on
image objects with an overall accuracy of 83.6 percent. Prob-
ability rules applied on accurate field boundaries increases
classification accuracy especially for agriculture land cover
mapping. Lucas
et al
. (2007) mapped agricultural land covers
with an overall accuracy of 84.9 percent using field boundar-
ies. Conrad
et al
. (2010) classified different irrigated crops
on per-field basis while attaining an overall accuracy of 80
percent.
Conclusion and Recommendations
Very high-resolution (
VHR
)
CIR
imagery showed great potential
for agriculture land cover classification. User defined rules
implemented on image objects have proved its effectiveness
to extract the desired land cover with higher level of ac-
curacy. The presented methodology helps to overcome the
traditional “salt and pepper” problem for
VHR
data and gives
the advanced user the flexibility to integrate expert knowl-
edge in the classifier. Good cadastral data is very useful for
agriculture land cover classification. However, accuracy of
field boundaries is of critical importance. It is assumed that
within one field, only one class exists. Field boundaries ought
to match with the crop dynamic boundaries for most of the
area. Otherwise, per-field classification be avoided or field
boundaries should be created/updated according to De-Wit
and Clevers (2004), Conrad
et al.
(2010), or Montaghi
et al.
(2013). Along with the advantages, developing rules based on
threshold values also has a downside. Too lenient rules tend
to include other crops in the desired class, and too strict rules
tend to miss significant portions of the desired class. Develop-
ment of rules takes considerable time and expert knowledge
of desired classes. Besides, these rules are implemented on
image objects created by segmentation. The segmentation
process of such high-resolution data takes an ample time and
requires substantial computer processing resources.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
August 2015
675
