CLASSIFICATION EFFICIENCY OF HIGH SPATIAL RESOLUTION SATELLITE IMAGERY 
IN THE STUDIES OF VEGETATION, FORESTRY AND HUMAN SETTLEMENTS 
M. Bajié, 
Student, Faculty of Forestry, Zagreb, Croatia 
ISPRS Commission VII 
KEYWORDS: High resolution, Quick Bird, Classification Efficiency, Human Settlement, Forestry, Fuzzy-C Means, Maximum 
Likelihood, Accuraccy 
ABSTRACT 
As an approach to better understanding of the expected benefits and shortcomings in application of high resolution satellite 
imagery in the studies of forestry, vegetation and human settlements, classification efficiency was analyzed on the imagery set: Im 
panchromatic, 3 m and 15 m color, 3m and 15 m IR, and color and IR merged and sharpened by the 1-m panchromatic image. 
The analysis is mainly focused on the evaluation of the classification efficiency for the forest, isolated trees, for the vegetated 
areas, buildings, streets and roads, by means of K-means, Fuzzy-C Means, ISODATA, One Pass Clustering, Maximum Likelihood 
and Minimum Distance to Mean methods. Relative value for the application was found out and briefly discussed. This work was 
made by TNTlite software system on a PC and imagery used was from the Earth Watch demo set. 
1. INTRODUCTION 
Announced commercial satellites should have spacial 
resolution within the area of several meters covering a few 
channels of visible and infra-red band (Fritz L.W., 1996). 
Great interest in. the expected benefits and shortcomings in 
application of the mentioned high resolution satellite 
imagery in the studies of forestry, vegetation and human 
settlements are justified since the resolution of Landsat and 
Spot is 30m, 20m, 10m. Therefore, classification efficiency 
was analyzed on the imagery set: 1m panchromatic, 3m and 
15m color, 3m and 15m IR, 4m color and IR, and color and 
IR merged and sharpened by the 1-m panchromatic image, 
such as the satellites manufactured by the company 
EarthWatch (Earth Watch, 1996 ) should be imaging. Since 
the link with the Early Bird satellite has not been 
established in the meantime, only the results referring to the 
Quick Bird satellite, which is to be launched, are used in the 
work. Analysis is made on the scene which includes a 
number of small houses, roads, pastures, water, forest, with 
the aim to find, by automatic classification, with or without 
supervision, how to make difference between artificial and 
natural areas applying the following classification methods: 
K-Means, Fuzzy-C Means, ISODATA, One Pass Clustering, 
Maximu Likelihood and Minimum Distance to Mean 
(Microlmages, 1997b). Since the ground-truth data are not 
accessible, selected samples include 1m panchromatic 
images (Earth Watch, 1996). Based on these samples, 
evaluation of the relative value of the imagery sources, 
classification methods for targeted application is made. 
Classification conducted on all the channel with selected 
methods on the examples of the analysis of objects. After 
the survey, only the results closest to the goal which was set 
have been selected. This work was made by TNTlite 
software system on a PC (Microlmages, 1997a). 
2. IMAGERY SOURCES, OBJECTS AND METHODS 
Images used in the work are taken from the following 5 
sources (Earth Watch, 1996) and were formatted at 
512 x 512 pixels due to the limitation of TNT Lite 
(Microlmages, 1997a): 
1. QuickBird Simulated Panchromatic 1 meter per pixel 
image greyscale, one band, 
2. QuickBird Simulated True Color Multispectral 4 meter 
per pixel image, Bandl, Blue, 450-520nm; Band 2, Green, 
520-600nm; Band 3, Red , 630-690nm. 
3. QuickBird Simulated Color IR Multispectral 4 meter per 
pixel image, Band 2, Green, 520-600nm; Band 3, Red , 630- 
690nm; Band 4, Near IR, 760-900nm. 
4. QuickBird Simulated Pan sharpened true color 
Multispectral Image, 3 bands. 
5. QuickBird Simulated Pan sharpened color IR 
Multispectral Image, 3 bands. 
Figures 2,3,4,5 are given in greyscale while the original 
ones are in color. Fig. 1 is used as ground-truth data. 
Figures were separated into channels before classification; 
R, G, IR are used with 4m resolution and R, G, B, IR with 
1m resolution. 
A number of classes which can be visually differentiated 
were selected from the Fig 1. The analysis should show 
which of them can be differentiated by automatic 
classification. The classes include: 
1. Forest - taller and shorter trees, isolated trees, clearance 
within the forest and shadow. 
2. Grass - larger areas, smaller areas in the forest and 
settlements around the roads. 
3. Water - at one place in a small pool. 
4. Houses - terraced and detached. 
5. Roads - outside and inside the settlement. 
Two classification methods are selected, supervised and 
unsupervised. At training data selection, samples from 
different locations were taken. Due to limited accuracy of 
the ground-truth data (1m resolution), the training data at 
supervised methods are not precise. That is why 
unsupervised methods have advantage since additional 
source of errors does not exist. Unsupervised methods have 
some parameters which can be determined but only some of 
them affect the results. Parameters which have the influence 
can be found at Simple One Pass Clustering 2, K-Means 4, 
Fuzzy-C Means 1, ISODATA 5. These parameters are 
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 657