mages
tona
nd for
y Was
ned on
asking
je crop
| some
of the
ure for
€ than
n, and
ication
vector
et al.
er and
ery is
ssified
is then
on the
> field.
porates
inly on
shown
ication
1e land
logical
sa and
sharma
cami ef
ing of
hrough
atellite
je was
lasses.
ype of
puting
Id and
y class.
ind the
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
2. THE STUDY AREA AND DATA DESCRIPTION
The study area is a 18 km by 13.5 km site situated near the
town of Karacabey, Bursa in northwest of Turkey (Figure 1).
The geographic boundaries of the area are N40?07'44" -
N40?13'43" and E28?10'31" - E28?20'28". The main crops
erown in the region are tomato, corn, pepper, wheat, onion,
and sugar beet. The villages that fall within the study area
include — Akhisar, Eskisaribey, Hotanli, — Ismetpasa,
Kucukkaraagac, Sultaniye, Yenisaribey, and Yolagzi. In the
region, a land consolidation project was performed between
1988 and 1992. Therefore, majority of the fields have regular
shapes that affect the classification accuracy. However, a
significant number of small fields exist in the area. The size
of the fields range from 0.0074 to 48 ha. The elevation
difference across the study area is very small (within 10 m).
The Landsat7 ETM- images (Path:180, Row:32) used in the
classification were acquired on May 15, 2000, July 2, 2000,
and August 19, 2000. All images were cloud free and of good
quality. A 600 pixels x 450 lines (multispectral) subscene
covering the study area was extracted to perform the
proposed sequential masking classification procedure.
green x
e
z3
rein i *Ankara
"x
- "
© la 9 Bursa
Study
Area
Figure 1. The study area.
3. DATA PRE-PROCESSING
The field boundaries were digitized from the cadastral map
sheets. The maps were converted into a raster form by
scanning them. Each raster map was registered to Gauss-
Kruger (Zone-5) projection and datum ED50 using six grid
intersections. The registration was based on first-degree
polynomial and nearest neighbor resampling techniques. The
registration accuracies were between + 1.14 m and + 3.59 m.
Next, the field boundaries were manually delineated through
on screen digitization and stored as vector polygons in the
database. Each polygon was assigned an identification
number and the crop types were recorded as attributes for
those fields from which reference information was collected
during site visit. There were a total of 2977 fields.
The multispectral bands (1-to-5 and 7) and the panchromatic
band were merged for each image date. The fused image
retains the spatial resolution of the panchromatic band, yet
provides the spectral properties of multispectral bands.
Furthermore, the field-based classification of a higher spatial
resolution image would be adventageous for small fields
since the number of pixels falling within the fields will
increase. The fused images were geometrically corrected to
Universal Transverse Mercator (UTM)-zone 35 projection
and the European Datum 1950 (ED50). The geometric
193
correction was based on second-degree polynomial and
nearest. neighbor resampling techniques. The root mean
square error (RMSE) values were computed as + 0.52 pixels
for the May image, + 0.67 pixels for the July image, and +
0.59 pixels for the August image.
We displayed the merged August image on the screen with
the digitized field boundaries overlaid, and those fields (275
in total) that contain multiple crops were identified through
visual inspection. The sub-boundaries within the fixed
geometry of these fields were then delineated through on
screen digitization and each sub-field was assigned an
identification number. There were 424 new sub-fields. With
these new sub-fields the total number of fields increased to
3401.
4. THE METHODOLOGY
A total of 1083 fields were visited on the ground. For each
field, crop type was collected as essential information and the
database was populated with the new data. For each image
date, the training samples were selected from all crops found
in the study area. The field boundaries were displayed on the
screen in superimposition with the raster imagery. For each
class, those fields from which reference information was
collected during site visit were selected through a database
query. Thereby, the boundary pixels were avoided. The
following classes were defined for the May image: bare soil,
wheat, clover, pasture, rice, pea, and onion. The July classes
include corn, residue, tomato, sugar beet, clover, pasture,
pepper, watermelon, bare soil, rice, cauliflower, and onion.
The classes defined for the August image are the same as the
July classes except for onion, which does not exist in August.
For each class, a group of representative pixels were
delineated within designated fields. Approximately 1096 of
the ground-visited fields were used in training, while the
other fields were set aside to be used as check fields to
perform accuracy assessement.
A per-pixel supervised classification of the images was
performed using the maximum likelihood method. For each
image date, the classification was carried out on bands |, 2, 3,
4, 5, and 7. In addition, Principal Component Analysis (PCA)
was carried out to obtain new channels. In the present case
the first 4 components of the May, July, and August images
provided 99.77%, 99.60%, and 99.76% respectively of the
total variance of the original data sets. Therefore, the first 4
PCs were used in the classification.
Upon completing the per-pixel classification, the integrated
analysis of the classified images and the vector field data was
carried out. To eliminate the effect of the boundary pixels on
the classification, a narrow corridor was generated along the
boundary inside each field. Field-based analysis was
performed by computing the class percentages within inside
corridor boundaries of each field and applying the entire field
the label of the modal class. For each field, the class
percentages and the final class label were automatically
inserted into the database. To compute classification
accuracy, the reference data were compared, field-by-field, to
each of the six classified images. The error matrix was used
to measure the accuracy of each classified output. The
accuracies of individual classes and the overall accuracy of
all bands and the first 4 PCs classifications are given in tables
| and 2.
|
Ez
EE
