International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
software. As the study is still in progress, the other dimensions
refinement using class merging. Image classification was
Internc
are being explored using integration of visual interpretation, performed using maximum likelihood algorithm. Prior to the SPEC
image processing, and GIS. classification execution, computation of statistical separability RT
between classes was done using transformed divergence and 2 Veo
4.2.1. Data and softwares Jeffrey-Matushita indices (Jensen, 1996). Thirdly, post-
classification using selective majority filtering was applied in
Two image dataset were used in this study, ie Landsat order to aggregate pixels of patchy classes into most common 3 Bar
Enhanced Thematic Mapper Plus (ETM+) bands 1-5 and 7, label within a given window, and to simultaneously preserve
and Quickbird high spatial resolution imagery with particular classes that are considered minority within a given
multispectral bands 1-4 and panchromatic. The ETM+ imagery window (e.g. linear features with 1-2 pixels width). By this 4 Paw
(path/row 120/065) was recorded on 21 August 2002, while the selective majority, a pixel-based generalisation can be applied sur
Quickbird imagery was recorded on 31 August 2002. The without losing important information conveyed by particular
whole area covered by the Quickbird is also covered by the individual pixels.
Landsat. In this study, two image processing softwares were
used, ie. ENVI 4.0 for most processing tasks, and ERDAS SPAT
Imagine 8.7 for particular ones. The ENVI software was 5. RESULTS AND DISCUSSION ] Wa
mainly used for making image subset, selecting samples 2 Veg
through regions of interest (ROIs), assessment of samples' 5.1. The Multidimensional Classification Scheme and
statistics, execution of multispectral classification and 3 Bar
assessment of classification accuracy. The ERDAS Imagine Figure 1 shows how particular categories under each dimension soil
8.7 was mainly used for multiresolution image merging, image are broken down into subtler classes. Based on the developed
reprojection and resampling, and recoding of pixel values categorisation, examples of spectral-related cover types 4 Bui
related to LC labels. classification using Landsat ETM+ and Quickbird imagery are sur
given.
4.2.2. Analysis
5.2. Example for the First Dimension: Automatic TEMI
During the first stage, each image dataset was treated Classification | Le
differently. After geometric correction and subset cropping, the rel
Landsat ETM+ data was prepared for multispectral Since the large number of classes obtained from the spectral 2 Ve
classification at 30 m pixel size. Meanwhile, a multi-resolution classification contains similar generic LC categories, a class rel
merging of Quickbird imagery using Brovey transform (Vrabel, merging operation needs to be run. During this stage, 40 3 Op
1996) was carried out in order to create a new colour composite spectral-related tentative cover classes from Landsat ETM+ rel
imagery with higher spatial resolution, i.e. 0.60 m. Meanwhile, image was merged to 27 LC classes with respect to the 4 De
The original Quickbird multispectral image dataset (2.4 m specified categories and spatial resolution under spectral rel
spatial resolution) was also preserved for multispectral dimension of the versatile LU classification scheme. By using
classification. the same procedure, 85 tentative classes obtained from
multispectral classification were merged to 48 spectral ECO
Image classification was performed in three stages. Firstly, dimension LC classes according to the versatile LU | Aq
ROI-based sampling that was performed interactively. classification scheme. Figure 8 shows the result.
Selection of ROIs was mainly based on the collected field data, 2 We
even though some additional ROIs were chosen based on local Accuracy assessment of the classified Landsat ETM+ and em
knowledge, topographic map as well as available aerial Quickbird images showed that the level of accuracy increases 3 Lo
photographs. ROI names were given with respect to the when the post-classification processes applied (Table 2). The lan
prepared classification scheme with a slight modification, c.g. immediate result of multispectral classification, i.e. original 4 Mo
shallow waterl, shallow water2, high density broadleaves on classified image was less accurate as compared to the classified lan
shaded areas. Every time a ROI is chosen, the sample statistics images followed by class merging. Class merging > Bui
were cvaluated and the class separability between existing ROIs consequently reduces the number of pixels of omission and
was also calculated. Especially for Quickbird image dataset, commission. This particularly gives positive effect for tentative
the ROIs selection was also guided by the display of Brovey- classes having relatively similar characteristics, e.g. shallow DIME
transformed multiresolution imagery. By doing so, water 1 and shallow water 2 , which were then be merged into I Wa
homogeneity within each ROI could be evaluated directly, both shallow water. :
visually and statistically. Secondly, image classification and = For
Table 2. Accuracy level of the classified images: original, merged, and majority-filtered classes. 3 Ag
Image data and | Accuracy levels of classified data with respect to Versatile LU Classification Scheme 4 Set
number of bands Original classified image Merged classes Global majority Selective majority inf
filtering filtering
Nr.of Accuracy Nr.of Accuracy Accuracy Accuracy
classes Overall & Kappa* | classes | Overall & Kappa" | Overall & Kappa" Overall & Kappa"
Landsat-7 ETM+ 40 86.84 % 27 92.56 % 94.38% 94.13%
(6 bands) (0.8628) (0.9211) (0.9434) (0.9378) LEVEI
Quickbird 85 68.75 % 48 79.02 96 87.05% 85.90% (>100
(4 bands) (0.6813) (0.7829) (0.8656) (0.8539)
Figur
* The Kappa coefficients are put within brackets
i9»!
Un
c
