International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
LARGE AREA LAND COVER CLASSIFICATION WITH LANDSAT 
ETM+ IMAGES BASED ON DECISION TREE 
Liang ZHAI, Jinping SUN, Huiyong SANG, Gang YANG, Yi JIA 
Key Laboratory of Geo-Informatics of NASG, Chinese Academy of Surveying and Mapping, Beijing, China 
zhailiang@casm.ac.cn 
Commission VII, WG VII/6: Remote Sensing Data Fusion 
KEY WORDS: land cover classification, decision tree, C5.0, MLC 
ABSTRACT: 
Traditional land classification techniques for large areas that use LANDSAT TM imagery are typically limited to 
the fixed spatial resolution of the sensors. For modeling habitat characteristics is often difficult when a study area 
is large and diverse and complete sampling of environmental variables is unrealistic. We also did some researches 
on this field, in this paper we firstly introduced the decision tree classification based on C5.0, and then introduced 
the classification workflow. The study results were compared with the Maximum Likelihood Classification result. 
Victoria of Australia was as the study area, the LANDSAT ETM+ images were used to classify. Experiments show 
that the decision tree classification method based on C5.0 is better. 
1. INTRODUCTION 
Detailed and accurate land cover data are widely 
used by various organizations, such as national, 
regional, local governments and private industries, as 
well as educational and research organizations because 
they are the basis for many environmental and 
socioeconomic applications (Perera and Tsuchiya, 
2009; Heinl et al., 2009). The suitability of remote 
sensing for acquiring land cover data has long been 
recognised and land cover mapping with using satellite 
data has received growing attention in the last 20 years., 
but the process of generating land cover information 
from satellite data is still far from being standardised or 
optimised (Foody, 2002; Lu and Weng, 2007; Heinl et 
al., 2009). 
Currently, particularly in times of global change, 
global land cover mapping has drawn much attention to 
many countries or organization. Till now, there are a 
number of global land cover products exist, such as 
IGBP DISCover, the MODIS land cover product, 
UMD land cover product, Global Land Cover 2000 
(GLC2000, Bartholomé & Belward, 2005) and 
GLOBCOVER (Loveland et al., 2000; Friedl et al., 
2002; Hansen et al., 2000; Herold et al 2008). These 
maps have been developed in response to the need for 
information about land cover and land cover dynamics. 
They all have been produced from optical, moderate 
resolution remote sensing and thematically focused on 
characterizing the different vegetation types worldwide 
(Herold et al 2008).Large area land cover 
classification still has many difficulties. 
There are three key problems of classification: 1) 
given a set of example records, 2) build an accurate 
model for each class based on the set of attributes, 3) 
use the model to classify future data for which the 
class labels are unknown. Common classification 
models are: neural networks, statistical models, 
Decision tree and genetic models. Decision tree has 
many great advantages in the remote sensing 
classification, which has been successfully used in 
many situations. Liu Zhongyang adopted the decision 
tree classification method based on LANDSAT TM 
image to present coverage situation of Zhengzhou city 
and proved that the decision tree classification method 
has obvious advantages, such as exact classification, 
efficient, definite classification criterion, intuitive 
classification structure controllable classification 
precision automated classification, etc (Liu 
Zhongyang, 2010). There are also many researches on 
C5.0, for example, this algorithm used in NLCD 2000 
to do land cover classification (Homer, Collin, 2000), 
to estimate tree canopy density (Chengquan Huang, 
2000), and so on, these applications all got good 
results. 
This paper introduced the C5.0 algorithm, and 
provided a C5.0 land cover classification platform, 
this platform also been used to classify the images of 
Victoria. Further, the classification result was been 
compared with MLC classification result, it proved 
that the C5.0 classification method is excellent. 
1.1 Study Area and Data 
We chose Victoria of Australia as our study Area. 
Victoria located in southeast of Australia, its location 
in Australia is as shown in figure i. Victoria is the 
smallest mainland state, and Australia’s second city of 
--- Melbourne is located in this state. Victoria's 
climate contains Mediterranean climate, temperate 
maritime climate, and some Savannah climate. The 
ecoregions that covered Victoria are Murray-Darling 
woodlands and mallee, Southeast Australia temperate 
savanna and Southeast Australia temperate forests. 
Various climate and terrain lead to rich land cover 
types in Victoria.