Comparison between pixel-based and object-oriented classification approaches in urban area of the arid environment

1
Comparison between pixel-based and
object-oriented classification approaches in
urban area of the arid environment

• Qian Jinga,b, Zhou Qiminga, Hou Quana

a Department of Geography, Hong Kong Baptist
University, Kowloon Tong, Kowloon, Hong
Kong b Xinjiang Institute of Ecology and
Geography, Chinese Academy of Sciences
2
Outline

• Introduction
• Study area and data
• Methods
• Results and discussion
• Conclusion

3
Introduction

• Urban development is one of the major forces
  causing environmental change in aridzone of
  China.
• In Xinjiang, expansion of urban areas is
  concentrated within limited space of oases, with
  constraints such as water resources.
• With the rapid increasing population, the
  expansion of built-up areas are accelerated in
  the past decades.

4
Introduction

• Delineating built-up areas from its background
  has been a constant challenge in remote sensing
  image processing (Erbek et al., 2004 Lo and
  Choi, 2004).
• With the increasing availability of
  high-resolution imagery, research has been
  focused on automated delineation of built-up
  areas using the images that have high frequency
  spatial variance with limited spectral
  resolution.
• Object-oriented approach has been developed for
  the segmentation of images for this.

5
The issues addressed

• In aridzone of China, the built-up areas are
  often surrounded by farmland.
• However, they may also confuse with nearby bare
  soil and stony desert, which present very similar
  spectral characteristics as construction
  materials such as concrete.
• The traditional pixel-based classification
  typically yield large uncertainty in the
  classification results.

6
Features in arid area
Built-up areas represented on the Landsat ETM
image show different types of cities with
significant spatial and spectral variations on
the images. Also notice the spectral similarity
between bare ground (river bed) and small cities
and settlements.
7
Objectives

• To find the most appropriate approach for
  auto-classification of built-up areas for the
  aridzone of China.
• Constraints
• Large areas to be covered in a short period time.
• Limited availability of high-resolution images.
• Cost-benefit concern
• Comparison between different classification
  approaches is addressed by this paper.

8
Study area and data
Landsat ETM image of the study area, acquired on
7 August 2000).
Map of the study area the Centre Town of Manas
County, City of Shihezi and part of regimental
farm of Division 8, at North Xinjiang Economic
Zone, China
9
Study area and data

• Centered at the city of Shihezi at north slope of
  Tianshan Mountain, Xinjiang Uygur Autonomous
  Region of China.
• At the centre of a large oasis.
• Three types of cities and settlements are
  identified
• Shihezi major city in the region, population
  200,000
• Manas county centre, population 20,000
• Liangzhouhu town settlement, population 5000
• Landsat ETM image is used, acquired on 7 August
  2000

10
Data pre-processing

• Geometric correction image-to-image registration
  using a geo-coded SPOT Pan image of 2002 as
  master image, with 37 Ground Control Points. The
  RMSE is less than 0.5 pixels
• Reference data interpreted from aerial photos
  acquired in 2000

11
Image classification approaches
Pixel-based Object-oriented
Philosophy Pixel as element Object as element
Information used Spectral only Spectral and spatial
Performance Expected to be superior, due to additional spatial information used
12
Methods tested

• Normalized Difference Built-up Index (NDBI)
• Maximum Likelihood Classifier (MLC)
• Object-oriented (O-O) image analysis

13
NDBI

• A pixel-based approach
• Using similar concept of vegetation index by
  delineating built-up areas from other background
  categories, only two bands of spectral data are
  used.
• The test classification attempts to delineate
  built-up areas barren soil, water-bodies and
  vegetation

NDBI (TM5-TM4) / (TM5TM4)
14
MLC

• A pixel-based approach
• Attempts to find clusters in the N-dimensional
  spectral space defined by N bands of spectral
  data
• ENVI/IDL was used for the test.

15
Classification scheme

• build-up area mixture of urban areas, settlement
  or lands under construction
• cropland cropland or fallow
• garden plot orchards, vineyards or nurseries
• sparse woodland low coverage mixture of shrub,
  desert scrub or bare ground
• dense woodland high coverage mixture of forest,
  shrub or shelter belt
• grassland pasture or desert grass
• river flat dry river bed or river flat
• water body reservoirs or fish ponds.

16
O-O Image Analysis

• Object-oriented approach
• Use both spectral and spatial information
• eCognition software was used for the test

17
O-O classification method
18
Accuracy assessment

• Stratified random sampling
• Referring to the ortho-corrected aerial photos.
• Totally 900 reference sites are selected as
  ground reference points.
• Error matrices were created for MLC and O-O
  results.

19
Results

• Classification results from different approaches
• The comparison of accuracy of different
  classification

20
Result NDBI

• The sparse woodland, bare ground and dry riverbed
  are merged into the same land-cover class as the
  background of built-up area.

21
Result MLC
22
Result O-O
23
Error matrix - MLC
Classified Data Reference Data Reference Data Reference Data Reference Data Reference Data Reference Data Reference Data Reference Data On map On ground Correct PA UA Conditional Kappa
Classified Data 1 2 3 4 5 6 7 8 On map On ground Correct PA UA Conditional Kappa
Water body 57 2 1 0 0 0 0 0 60 91 57 62.64 95.00 0.9444
Bottomland 8 95 3 1 0 0 0 0 107 103 95 92.23 88.79 0.8734
Build-up area 4 5 85 11 12 3 4 0 124 117 85 72.65 68.55 0.6385
Sparse woodland 2 0 5 92 14 4 23 1 141 109 92 84.40 65.25 0.6046
Cropland 0 0 2 1 126 3 0 39 171 189 126 66.67 73.68 0.6669
Garden 0 0 0 0 0 49 0 0 49 90 49 54.44 100.00 1.0000
Grassland 0 0 1 4 0 2 59 0 66 86 59 68.60 89.39 0.8827
Dense woodland 20 1 20 0 37 29 0 75 182 115 75 65.22 41.21 0.3260
Column Total 91 103 117 109 189 90 86 115 900 900 638
Overall Classification Accuracy 70.89 Overall
Kappa Statistics 0.6633
24
Error matrix O-O
Classified Data Reference Data Reference Data Reference Data Reference Data Reference Data Reference Data Reference Data Reference Data On map On ground Correct PA UA Conditional Kappa
Classified Data 1 2 3 4 5 6 7 8 On map On ground Correct PA UA Conditional Kappa
Water body 95 4 3 0 1 1 1 0 105 99 95 95.96 90.48 0.8930
Bottomland 0 100 0 0 0 0 0 0 100 107 100 93.46 100.00 1.0000
Build-up area 2 0 89 0 3 1 17 5 117 105 89 84.76 76.07 0.7291
Sparse woodland 0 1 3 104 1 0 0 0 109 107 104 97.20 95.41 0.9479
Cropland 0 0 2 1 0 1 152 2 158 198 152 76.77 96.20 0.9513
Garden 0 2 1 0 2 0 12 85 102 95 85 89.47 83.33 0.8137
Grassland 0 0 7 2 0 88 3 0 100 91 88 96.70 88.00 0.8665
Dense woodland 2 0 0 0 91 0 13 3 109 98 91 92.86 83.49 0.8147
Column Total 99 107 105 107 98 91 198 95 900 900 804
Overall Classification Accuracy 89.33 Overall
Kappa Statistics 0.8773
25
Discussion Comparison between MLC O-O
MLC O-O
Overall accuracy 70.9 89.3
Kappa 0.663 0.878
PA of built-up area 72.7 84.8
UA of built-up area 68.6 76.1
26
Discussion usability of NDBI

• The NDBI method is found to be unable to
  differentiate urban areas from the background
  features such as sparse woodland, bare ground and
  dry riverbed in arid regions.
• The usability of such a pixel-based spectral
  classifier is severely limited in the arid
  regions mainly due to the common presence of
  land-covers of bare ground and dry riverbed,
  which have similar spectral response with
  built-up areas.

27
Discussion MLC versus O-O

• The object-oriented classifier yields
  significantly better overall accuracy than the
  MLC method.
• For built-up areas, however, the difference
  between MLC and O-O methods is less significant.
• Both methods appears to have less omission errors
  but larger commission errors.

28
Discussion shortfalls of the O-O method

• The classification accuracy depends on the
  quality of image segmentation. If objects are
  extracted inaccurately, subsequent classification
  accuracy will not improve.
• Classification error could be accumulated due to
  the error in both image segmentation and
  classification process.
• Once an object is misclassified, all pixels in
  this object will be misclassified.

29
Conclusion

• This study has compared classifiers regarding to
  built-up area delineation in aridzone of China.
• Although the overall accuracy of the O-O approach
  is significantly better than that of MLC, there
  is less significant difference for the built-up
  area class.
• Further research will be focused on the impact of
  spatial resolution of images and the efficiency
  of different classifiers.

30
Thanks