2004
—
band
5. All
> full
y, (he
ately
ected
g an
icular
ive of
in the
crop
valid)
round
ining
vey in
d for
n, the
n the
parcel
es) to
s hard
| Pasa.
)-scale
zitizing
()-scale
tabase;
satellite
)S wert
control
national
n using
roduced
y, since
magery.
pology.
jatching
cultural
mdaries,
nap and
system.
with the
- further
indsat 7
h the 15-
d bands
spectral
nds. The
ace. It is
in lower
nd in the
alike the
International Archives of the Photogrammetry, Remote Sensing
popular technique of IHS type enhancement, in which the
spectral characteristic of each bands are not preserved (Liu,
1999). The Atlantic technique has been integrated to PCI
Geomatica and available through the function IMGFUSE (PCT,
2003). The resulting output images were then used as input to
subsequent geometric correction process. It is important to note
that the geometric correction is not required prior to data fusion
in case of Landsat 7 ETM+ since both data are co-registered at
the satellite.
Geometric correction: To integrate Landsat 7 ETM+ images
with the vector polygon parcel boundaries, it was necessary to
register both data sets to a common map coordinate system. For
this purpose, vector to image registration is performed using
Landsat 7 (L1G) images and vector parcel data as reference
layer. Georeferencing was carried out using just about 35
ground control points (GCPs) for each subscene through a
polynomial transformation and nearest neighbor resampling
method. For all images, RMSE values were calculated less than
0.5 pixels (7.5 m) in both X and Y directions.
Parcel based Analysis
The principal objective of this study was to develop a parcel-
based classification methodology in the integration of remote
sensing and GIS in order to map the summer (August) crops
within the agricultural land parcels. For this purpose, a number
of agricultural land cover classes were determined for each of
the Landsat 7 ETM+ images with the help of reference data
kept in a GIS database. Prior to collecting the training areas,
boundary pixels of adjacent agricultural parcels were excluded
using a binary raster mask. Since, boundary pixels manifest the
spectral mixture of two or more land cover classes, thus cause
misclassification. Therefore, only those pixels that fall within
the agricultural parcel were kept for further analyses.
For each scene, Bands 3, 2, 1 and 4, 5, 3 were utilized as RGB
components on separate video display windows. Displaying
bands 4, 5 and 3 as infrared colour composite is useful for
monitoring the vegetation development during training process.
The images were overlaid with vector polygon layer and then
through querying GIS database, boundary of each land cover
lype were separately displayed over the images to locate the
known land cover classes and define training areas. Block
training was carried out to select blocks of pixels from the
centre parts of the referenced land parcels. During training
procedure, heterogeneous land cover classes were avoided since
they would increase the spectral overlap between the classes. A
number of samples were collected for each class with respect to
amount and the size of the parcels as well as their dispersion
throughout the study area.
The August training set comprises the following eleven
information classes: corn (Cor), residue (Res), tomato (Tom),
Sugar beet (Sbe), clover (Clo), pasture (Pas), pepper (Pep),
Watermelon (Wme), uncultivated land/bare soil (Unc/Bso), rice
(Ric) and cauliflower (Cfl). The training samples were refined
through examining their histograms and scatter plots. Those
Classes that include two or more different spectral
Characteristics were subdivided into separate spectral classes.
Such as the corn subclasses: cor-01, cor-02, cor-03 and cor-04.
The classification accuracy is generally improved when each
Subclass represented as separate spectral class (Lillesand &
Kiefer, 1994).
and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
After completing the training selection for August image, the
same agricultural parcels were utilized to collect the training
samples for July and May images, respectively. However,
training areas were re-edited to reflect any variation in spectral
characteristic of a particular class or altering land cover type for
a particular scene. A final set of statistics were generated for
seven classes in May and twelve classes in July.
Optimum band selection: Once the training statistics were
assembled from each band for each land cover, a decision must
be made to determine the most suitable bands for discriminating
the land cover classes in the imagery. For this purpose, the
Principal Component Analysis (PCA) was performed to obtain
new channels in the classification process. The first four PCA
channels of each date having total variance higher than 99%
were used within the analysis. In addition to PCA channels,
most effective bands were identified through examining class
separability based on the divergence of the class signatures. As
a result, a subset of best four bands was extracted for each
scene. These were ETM+ Bands 3, 4, 5, 7 for May and July
scenes, and Bands 2, 3, 4 and 7 for August scene.
Classification: For each date, maximum likelihood classifier
was utilized with equal probability assumption for all classes.
The classification was essentially performed using the bands 1
to 5 and 7. In addition, the first four PCA channels and the best
four band combinations were also classified. Each classified
image was subsequently processed to aggregate the spectral
sub-classes into associated information classes using a scripting
language. The resulting classified images were then integrated
with the vector polygon layer to perform parcel-based analysis.
Parcel-based analysis was performed by computing the class
frequencies within each parcel and assigning the parcel the
class labels based on the highest frequency computed from
pixel-based classified image. The results of parcel-based
analysis were stored as attribute within the database.
Regional masking: Prior to the multi-temporal analysis, we
have performed a masking based on the knowledge obtained
from the relations between land cover classes and the
agricultural parcel boundaries. In our case, the knowledge
represents the cultivation practice in the study area. That is, the
cultivation of sugar beet was restricted by the government
agencies into a group of agricultural parcels. And similarly,
cauliflower was merely grown within the boundaries of Hotanli
village. Knowing these facts, the database was populated to
reflect the relations regarding to restriction zones for
cultivation. Using the database, regions of cultivation were
defined as regional masks to include and/or to exclude training
areas of sugar beet and cauliflower. Then using the regional
masks, parcel-based classification was performed on both July
and August images, where both sugar beet and cauliflower were
cultivated.
Multitemprol Analysis
The methodology presented in this part of the paper is based on
enhancing the discrimination potential of supervised image
classification through combining maximum likelihood
classifier, parcel-based analyses, and multi-temporal masking
technique (Figure 3). The method was implemented using
Landsat 7 ETM+ images. In addition, Principal Component
Analysis (PCA) channels and the best four bands were used in
the analyses. Through analyses of multi-temporal images set, it
is possible to achieve better discrimination of land cover classes
1087