Nguyen Dinh, Duong 
LAND COVER CATEGORY DEFINITION BY IMAGE INVARIANTS FOR AUTOMATED 
CLASSIFICATION 
Nguyen Dinh Duong 
Environmental Remote Sensing Laboratory 
Institute of Geography 
Hoang Quoc Viet Rd., Cau Giay, Hanoi, Vietnam 
duong.nd@hn.vnn.vn 
Technical Commission TC VII-5 
KEY WORDS: Automation, Classification, Remote Sensing, Land use/Land cover, Global change. 
ABSTRACT 
The definition of conventional land cover categories (legend) is usually formulated by statements which describe a land 
cover category by criteria derived from an interdisciplinary approach. Such a system of definition is suitable for 
conventional land cover mapping by ground observation or visual image interpretation when the interpreter combines 
image and auxiliary information to classify an object. Supervised image classification relies on statistical parameters of 
a class generated during training sampling which is in general nontransferable from one image to the other. 
Unsupervised classification with a clustering technique provides automated grouping, but there is no way to establish a 
fixed relation between a cluster code and a certain land cover category. Moreover the post-classification interpretation 
of results is time consuming and a subjective process that requires extensive ground truth data collection. Recently, 
research into automated classification of land cover was initiated by the author in the framework of the NASDA 
Research Announcement for the ADEOS-II satellite, which will carry a GLI sensor that has 6 spectral channels similar 
to those of LANDSAT TM. One of the issues of automated classification of GLI data is to develop a system for 
defining land cover in the image domain. This means a land cover category should be described by values derived from 
image data. The author has discovered several image invariants based on graphical analysis of the spectral reflectance 
curve of a pixel. The invariant found include modulation of the spectral reflectance curve, total reflected radiance index 
(TRRI) and spectral angles. These invariants seem to be quite stable for image data generated by the same sensor. In the 
paper the author reports on efforts to develop a digital definition of land cover using image invariants for automated 
classification. The proposed legend has been applied to a data set simulated for the future GLI sensor and the obtained 
classification result has proved the chosen approach to be correct. 
1 INTRODUCTION 
Land use and land cover change play a pivotal role in global environmental change. They contribute significantly to 
earth-atmosphere interactions and biodiversity loss, are major factors in sustainable development and human responses 
to global change, and are important for integrated modelling and assessment of environmental issues in general (IGBP 
Report No. 35 Land-Use and Land-Cover Change Science/Research Plan). This awareness has recently led to land use 
and land cover mapping activities at national, regional and global scales. One of the issues of land cover mapping is to 
standardise classification schemes from both technical and typological points of view. In conventional land cover 
mapping, each land cover unit is usually defined in the form of a statement that describes the land cover category by a 
set of attributes derived from an interdisciplinary approach (vegetation and soil sciences). Such a system of definition is 
suitable for conventional land cover mapping by ground observation (field survey) or visual image interpretation when 
the interpreter combines image and auxiliary information to classify an object. In computer processing there are two 
methods of analysis which are mostly used for land use/land cover classification. They are supervised and unsupervised 
classification. Supervised image classification relies on statistical parameters of a class generated during training 
sampling which are in general nontransferable from one image to another. Unsupervised classification with a clustering 
technique provides automated grouping but there is no way to establish a fixed relation between a cluster code and 
certain land cover category. Moreover the post-classification interpretation of results is time consuming and a subjective 
process that requires extensive ground truth data collection. The huge amount of imagery information collected by high 
resolution remote sensing satellites such as LANDSAT and SPOT and especially by the medium resolution multi-sensor 
satellites TERRA (1999) and ADEOS-II (2001), which will complete global coverage every 4 days, requires 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 985