Home Technology Aerial imaging Evaluation of conventional digital camera scenes for Thematic Information Extraction

Evaluation of conventional digital camera scenes for Thematic Information Extraction

Evaluation of conventional digital camera scenes for Thematic Information Extraction

H. S. Lim


H. S. Lim

M. Z. MatJafri and K. Abdullah
School of Physics
Universiti Sains Malaysia,
11800 Penang

Introduction
The increasing availability of remote-sensing images, acquired periodically by satellite
sensors on the same geographical area, makes it extremely interesting to develop the
monitoring systems capable of automatically producing and regularly updating landcover
maps of the considered site (Bruzzone, et al., 2002). Airborne remote sensing was
selected in this present study because of several reasons. First was the airborne images
can provides higher spatial resolution for mapping a small study area. Second was the
airborne data acquisition can be carried out according to our planned surveys. It’s not like
the satellite data was fixed on time of satellite overpass the study area only. Third, for
airborne remote sensing, atmospheric correction was not need to apply to the analysis
data because atmospheric correction only improves R2 and RMS significantly for
turbidity, this is an advantage since one step of the retrieval process can be eliminated
(Koponen, et al., 2002). The objective of this study is to investigate the potentiality of
using digital camera imagery for land cover mapping. In this study, images captured from
a digital camera were used for land cover mapping. Supervised classification methods
were applied to the digital images. Many researchers used the Maximum Likelihood
method in their study (Donoghue and Mironnet, 2002). The monitoring task can be
accomplished by supervised classification techniques, which have proven to be effective
categorization tools (Bruzzone, et al., 2002). Accuracy assessment also has been done in
this study.

Study Area



Source: Microsoft Corp., 2001.
Figure 1. The study area

Merbok River estuary was chosen as the study area in this study. The study area is
located at latitude 5° 39′ N to 5° 41′ N and longitude 100o 20′ E to 100o 24′ E (Figure 1).
The images were captured from a light aircraft flying at an altitude of 8000 feet on 9
March 2002.

Evaluation of conventional digital camera scenes for Thematic Information Extraction

Data Analysis And Result

The size of the airborne colour digital images used in this study of the Merbok River
estuary, Kedah is 1200 pixels by 1792 lines namely Area A, Area B and Area C (Figure
2). Three supervised classification methods were performed to the digital images
(Maximum Likelihood, Minimum Distance-to-Mean, and Parallelepiped). Training sites
were needed for supervised classification and selected based on the colour in present
study. The digital image was classified into 4 classes, such as water, forest, land and
urban. Accuracy assessment was done in this study to compute the probability of error for
the classified map. A total of 200 samples were chosen randomly for the accuracy
assessment. Many methods of accuracy assessment have been discussed in remote
sensing literatures. Three measures of accuracy were tested in this study, namely overall
accuracy, error matrix and Kappa coefficient. In thematic mapping from remotely sensed
data, the term accuracy is used typically to express the degree of ‘correctness’ of a map
or classification (Foody, 2002). Figure 3 shows the flow chart for data processing of the
images.



Figure 2: Digital images used in image classification



Figure 2: Digital images used in image classification



Figure 3: Flow chart for data processing of the images

Evaluation of conventional digital camera scenes for Thematic Information Extraction

(A) Area A
Kappa coefficient and overall accuracy results of the three measures of accuracy are
shown in Table 1. The overall accuracy is expressed as a percentage of the test-pixels
successfully assigned to the correct classes. The results obtained are presented in Tables
1, 2 and 3, where the overall classification accuracy, the confusion matrix and the
accuracy of each class using Maximum Likelihood, minimum distance-to-mean and
parallelepiped classification are given, respectively. From the present analysis, one can
see that the Maximum Likelihood classifier produced the best image classification
accuracy with the highest overall accuracy and Kappa coefficient. The overall
classification accuracies achieved by the proposed Maximum Likelihood classifier on the
digital image is 92.00 %. This followed by the Minimum Distance-to-Mean with the
overall classification accuracy of 85.50%, and Parallelepiped resulted in the overall
classification accuracy of 67.00%. A classified image using Maximum Likelihood
classifier is shown in Figure 4.

Table 1: The overall classification accuracy and Kappa coefficient

Classification method
Overall classification accuracy (%)
Kappa coefficient

Maximum Likelihood
92.00
0.884

Minimum Distance-to-Mean
88.50
0.832

Parallelepiped
67.000
0.561

Data
Supervised classification
was performed to the
digital images
Some samples training
sites were choose
Accuracy Assessment

Table 2: The confusion matrix results

Classified Data
Reference Data

Forest
Water
Water Turbid
Land
Total

Forest
72
2
0
1
75

Water
3
65
3
1
72

Turbid Water
0
4
33
1
38

Land
0
0
0
14
15

Total
76
71
36
17
200

Table 3: The accuracy of each class using Maximum Likelihood classification.

Class
Maximum Likelihood

Producer Accuracy (%)
User Accuracy (%)

Forest
94.737
96.000

Water
91.549
90.278

Turbid Water
91.667
86.842

Urban
82.353
93.333

Evaluation of conventional digital camera scenes for Thematic Information Extraction



Figure 4: The classified image obtained using Maximum Likelihood classifier for
Merbok River estuary (Green = Forest, Blue = Water, Orange = Land, and Light Blue =
Turbid Water)

(B) Area B
The Kappa coefficient and overall accuracy value for the three-classification technique
are shown in Table 4. The others accuracy assessment results are presented in Tables 5
and 6, where the Kappa coefficient, the confusion matrix and the accuracy of each class
using Maximum Likelihood, minimum distance-to-mean and parallelepiped classification
are given, respectively. From the present analysis, one can see that the Maximum
Likelihood classifier produced the best image classification accuracy with the highest
overall accuracy and Kappa coefficient. The overall classification accuracies achieved by
the proposed Maximum Likelihood classifier on the digital image is 95.00 %. This
followed by the Minimum Distance-to-Mean with the overall classification accuracy of
73.00%, and Parallelepiped resulted in the overall classification accuracy of 1.00%. A
classified image using Maximum Likelihood classifier is shown in Figure 5.

Table 4: The overall classification accuracy and Kappa coefficient

Classification method
Overall classification accuracy (%)
Kappa coefficient

Maximum Likelihood
95.000
0.866

Minimum Distance-to-Mean
73.000
0.457

Parallelepiped
1.000
0.008

Table 5: The confusion matrix results

Classified Data
Reference Data

Grass
Water
Land
Urban
Total

Grass
21
1
0
0
22

Water
0
154
1
0
155

Land
2
2
12
4
20

Urban
0
0
0
3
3

Total
23
157
13
7
200

Table 6: The accuracy of each class using Maximum Likelihood classification.

Class
Maximum Likelihood

Producer Accuracy (%)

User Accuracy (%)

Grass
91.304
95.455

Water
98.089
99.355

Land
92.308
60.000

Urban
42.857
100.000

Evaluation of conventional digital camera scenes for Thematic Information Extraction



Figure 5: The classified image obtained using Maximum Likelihood classifier for
Merbok River estuary (Green = Forest, Blue = Water, Orange = Land, and Red =Urban)

(C) Area C
The Kappa coefficient and overall accuracy value for the three-classification technique
are shown in Table 7. The others accuracy assessment results are presented in Tables 8
and 9, where the Kappa coefficient, the confusion matrix and the accuracy of each class
using Maximum Likelihood, minimum distance-to-mean and parallelepiped classification
are given, respectively. From the present analysis, one can see that the Maximum
Likelihood classifier produced the best image classification accuracy with the highest
overall accuracy and Kappa coefficient. The overall classification accuracies achieved by
the proposed Maximum Likelihood classifier on the digital image is 79.50 %. This
followed by the Minimum Distance-to-Mean with the overall classification accuracy of
76.50%, and Parallelepiped resulted in the overall classification accuracy of 10.00%. A
classified image using Maximum Likelihood classifier is shown in Figure 6.



Figure 6: The classified image obtained using Maximum Likelihood classifier for
Merbok River estuary (Green = Forest, Blue = Water, Orange = Land, and red =
Urban)

Table 7: The overall classification accuracy and Kappa coefficient

Classification method
Overall classification accuracy (%)
Kappa coefficient

Maximum Likelihood
79.50
0.70

Minimum Distance-to-Mean
76.50
0.652

Parallelepiped
10.00
0.069

Table 8: The confusion matrix results

Classified Data
Reference Data

Grass
Water
Land
Urban
Total

Grass
59
2
3
1
65

Water
5
72
10
2
89

Land
1
1
22
3
27

Urban
2
2
9
6
19

Total
67
77
44
12
200

Evaluation of conventional digital camera scenes for Thematic Information Extraction

Table 9: The accuracy of each class using Maximum Likelihood classification.

Class
Maximum Likelihood

Producer Accuracy (%)
User Accuracy (%)

Grass
88.060
90.769

Water
93.506
80.899

Land
50.000
81.481

Urban
50.000
31.579

Conclusion
In this study, Maximum Likelihood was the best classifier to extract thematic information
from remote sensed imagery. The high spatial resolution images gave a more detail
deposition mapping of the classified map. So it is good for a small coverage of study
area. From the result of the accuracy assessment, we were quite confident of the
classified shown. Digital camera imagery provides a cheaper way to acquired remote
sensed imagery for land cover mapping.

Acknowledgement
This project was carried out using the Malaysian Government IRPA grant no. 08-02-05-
6011 and USM short term grant FPP2001/130. We would like to thank the technical staff
and research officers who participated in this project. Thanks are extended USM for
support and encouragement.

Reference

  • Bruzzone, L., Cossu, R. and Vernazza, G. (2002). Combining parametric and nonparametric
    algorithms for a partially unsupervised classification of
    multitemporal remote-sensing images. Information Fusion 3, 289 -297.
  • Donoghue, D. N. M. and Mironnet, N. (2002). Development of an integrated
    geographical information system prototype for coastal habitat monitoring.
    Computers and Geosciences, 28, 129-141.
  • Foody, G. M. 2002. Status of land cover classification accuracy assessment. Remote
    Sensing and Environment, 80, 185-201.
  • Koponen, S., Pulliainen, J., Kallio, K. and Hallikainen, M. (2002). Lake water quality
    classification with airborne hyperspectral spectrometer and simulated MERIS
    data. Remote Sensing of Environment 79, 51- 59.
  • Microsoft Corp., Map of Kedah, Malaysia. (2001). [online].