Sequential masking classification of multi-temporal Landsat7 ETM +
images for field-based crop mapping in Karacabey, Turkey
M. TURKER* and M. ARIKAN
Middle East Technical University, Graduate School of Natural and Applied Sciences,
Geodetic and Geographic Information Technologies, 06531 Ankara, Turkey
(Received 5 January 2005; in final form 3 March 2005 )
Three Landsat7 ETM + images acquired in May, July and August during the
2000 crop growing season were used for field-based mapping of summer crops in
Karacabey, Turkey. First, the classification of each image date was performed on
a standard per pixel basis. The results of per pixel classification were integrated
with digital agricultural field boundaries and a crop type was determined for each
field based on the modal class calculated within the field. The classification
accuracy was computed by comparing the reference data, field-by-field, to each
classified image. The individual crop accuracies were examined on each classified
data and those crops whose accuracy exceeds a preset threshold level were
determined. A sequential masking classification procedure was then performed
using the three image dates, excluding after each classification the class properly
classified. The final classified data were analysed on a field basis to assign each
field a class label. An immediate update of the database was provided by directly
entering the results of the analysis into the database. The sequential masking
procedure for field-based crop mapping improved the overall accuracies of the
classifications of the July and August images alone by more than 10%.
1. Introduction
Remotely sensed data acquired by the operational satellites are more and more
widely used for crop mapping at regional or global scales. The availability of digital
multispectral images and the advances in digital processing and analysis have
enabled research scientists to have information about the type, condition, area, and
the growth of the crops. One crucial technique in detecting the crops from remotely
sensed data is the automated image classification. Most current automatic
classification techniques to obtain crop maps from digital imagery operate on a
per-pixel basis in isolation from other pertinent information. Therefore, per-pixel
techniques often yield results with limited reliability on areas where parcel size is too
small. The reliability of image classification can be improved by including a priori
knowledge about the contextual relationships of the pixels in the classification
process. Agricultural field boundaries integrated with remotely sensed data divide
the image into homogeneous units of image pixels. For each field, the geometry of
the boundaries defines the spatial context between the pixels contained within, and
enables those pixels to be processed in coherence. A final decision on the class
assignment of pixels within each field is taken based on the coherent processing of
these pixels. This is unlike per-pixel classification where the decision for each pixel is
*Corresponding author. Email: mturker@metu.edu.tr
International Journal of Remote Sensing
Vol. 26, No. 17, 10 September 2005, 3813–3830
International Journal of Remote Sensing
ISSN 0143-1161 print/ISSN 1366-5901 online # 2005 Taylor & Francis
http://www.tandf.co.uk/journals
DOI: 10.1080/01431160500166391

reached independently. Therefore, the standard per-pixel image classification can be
replaced by a classification which operates on a field basis.
Field-based image classification has been adopted by several researchers to obtain
land cover information from remotely sensed data. One of the earliest statements
was made by Derenyi (1979) 26 years ago, but GIS technology at the time was not
sufficiently advanced to implement it. After facing the problems of misclassification
caused by the per-pixel approach, Baker and Drummond (1984) suggested
employing a more useful and less noisy classification by means of a parcel-based
approach. Wooding (1984) segmented a SAR image into fields using field
boundaries and each field was treated as a separate unit for backscatter
measurement and image classification. Catlow et al. (1984) superimposed vector
data on per-pixel classified Landsat MSS image and visually assessed, on a per-field
basis, the accuracy of per-pixel classification. Janssen et al. (1990) integrated
topographical data stored in a GIS into an object-based classification. After
performing a per-pixel classification, a label was determined, for each object, based
on the modal class. Further, Janssen et al. (1992) detected the edges of subdivisions
separating different crops within existing field boundaries using edge detection. The
classification process and the labelling of each field was then performed in a similar
way as described in Janssen et al. (1990).
Aplin et al. (1999) developed a set of classification techniques to detect land cover
on a per-parcel basis from high resolution imagery. Of the techniques developed, the
per-field texture filtered classification was reported to provide the best results. Two
field-based image analysis schemes were developed by Turker and Derenyi (2000)
for monitoring the changes in land cover conditions within existing land use
boundaries stored in a GIS. In the first scheme, a modified parallelepiped image
classification was performed, one theme at a time. The results were then assessed on
a field basis to separate those fields which deviate from the a priori expectations. In
the second scheme, only image statistics were generated field-by-field to identify
those fields where significant changes occurred. Aplin and Atkinson (2001)
developed a method for classifying land cover at the sub-pixel scale based on pixel
segmentation for subsequent per-field classification. Smith and Fuller (2001) created
a parcel-based land cover map for Jersey. Their approach used vector land parcel
boundaries to subdivide images, aggregate the raster data within the land parcel and
classify on a per-parcel basis. De Wit and Clevers (2004) created a crop map of the
Netherlands by integrating multi-sensor satellite imagery, statistical data on crop
area, and parcel boundaries from a digital topographic map. The crop type was
determined, for each field, from the spectral and phenological properties of the field.
Lloyd et al. (2004) classified the land cover of a Mediterranean region within an
artificial neural network on a per-field basis. In addition to spectral information,
geostatistical and texture measures extracted from the co-occurrence matrix were
utilized. Aplin and Atkinson (2004) developed a technique for predicting missing
field boundaries to increase the accuracy of per-field classification. The technique is
based on the comparison of the within-field modal land cover proportion and local
variance, which provides an indication of the missing boundaries.
The field-based image classification techniques are commonly carried out by
integrating remotely sensed imagery and vector field boundary data (Brisco et al.
1989, Janssen et al. 1990, 1992, Aplin et al. 1999, Turker and Derenyi 2000, Aplin
and Atkinson 2001, De Wit and Clevers 2004). The integration between the two
datasets can be achieved: (i) before classification, (ii) during classification, and (iii)
3814 M. Turker and M. Arikan