Nguyen Dinh, Duong 
  
  
  
  
  
  
31 Muddy surface Muddy land 
21 Turbid water Turbid water 
1 Water and other hydrographic | Clear water 
bodies : 
12 Snow and ice 
  
  
  
  
Table 2. Proposal of classification system for single date remote sensing data 
3 DEFINITION OF LAND COVER CATEGORIES BY IMAGE INVARIANTS 
Each land cover category features unique spectral absorption characteristics. For a remote sensing sensor with a fixed 
spectral channel composition, these characteristics should be unique and stable for a certain land cover class. There are 
many ways to extract a feature which is unique for a certain land cover class, the author has chosen a method called 
Graphical Analysis of Spectral reflectance Curve (GASC) to define image invariants. Each land cover class can be 
described by a set of invariants derived from normalized pixel vector (Nguyen Dinh Duong, 1997 and 1998). According 
to the latest research result, each land cover class could be described by some of the following invariants: 
- Spectral curve modulation 
- Total reflected radiance index TRRI 
- Band ratios 
- Hue angle 
- Saturation angle 
- Difference of normalized spectral values 
The Hue and Saturation angles are computed based on a compression model of 6 spectral channel data (TM or GLI) 
into three components developed by the author using hexacone colour space. 
Table 3 gives an example of the digital definition for some land cover categories based on proposed image invariants. 
4 PRELIMINARY RESULT OF AUTOMATED CLASSIFICATION 
The above proposed land cover classification has been used for classification of TM and GLI simulated data. A 
computer program for automated classification has been written by the author. The program runs on DOS prompt in 
command line mode which provides batch processing ability. The control file contains beside basic information about 
the data set such as number of lines, rows, file names for different spectral channel data file etc. also a table of 
classification rules for land cover categories. Structure of digital legend for a land cover category is as follows: 
- Classification method (Graphical Analysis of Spectral Reflectance Curve) 
- Data set name (GLI 250m channels) 
- Number of classes (255 is maximum) 
The following is repeated for the number of classes 
- . Class code (between 1 and 255) 
- . Full name of class (127 characters) 
- Short name of class (6 characters) 
- RGB Colour for visualization (example 1 213 255) 
- M code - Modulation invariant 0 - 26 
- Dij min max - Difference of channels i and j (example D15 15 60) 
- T min max - TRRI (example T 2 10) 
- H min max - Hue angle between 0 - 360 (example H 15 150) 
- S min max - Saturation angle between 0 - 60 (example S 5 30) 
-  Pimin max - Normalized pixel value of channel i 0 - 100 (example PS 0 15) 
- A Aij min max - Absolute values of difference of channel i and j (example A15 0 20) 
-  Rij min max - Ratio of bands i and j (example R34 15.5 20.0) 
- END the end of the description for one class 
The threshold values for each invariant have been computed based on the normalized pixel vector. The normalization 
should be done so that it eliminates impact of the seasonal variability of solar radiation, sensor sensitivity and 
degradation and the quantization level of the sensor. 
  
988 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 
11 
  
The 
defin