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 
2) Tree Pruning Phase 
This phase should remove dependency on 
statistical noise or variation that may be particular 
only to the training set. 
Decision tree algorithm is a data mining induction 
techniques that recursively partitions a data set of 
records using depth-first greedy approach (Hunts et al, 
1966) or breadth-first approach (Shafer et al, 1996) 
until all the data items belong to a particular class. 
The decision tree modeling comes in two main 
branches: Breiman's Classification and Regression 
Trees (CART) and Quinlan's See5/C5.0 (and its 
predecessors, C4.5 and ID3). This study we mainly 
introduce the C5.0 model. 
2.1.2 C5.0 algorithm 
C5.0 is one of the most classic algorithms in 
decision tree models, which increased boosting 
technology on the basis of C4.5. According to C5.0 
algorithm, the original training sample set is 
considered the root node of the decision tree, and then 
the gainratio of every feature attribute are calculated. 
Some definitions were put forward: 
Information entropy: suppose S is the set of n data 
samples. The category attribute C has m different 
values, and it divides the sample set into m different 
category C;(i = 12,.,m). Suppose n, is the 
amount that the samples belong to C; in S. then the 
information entropy E(s) of S is defined as, 
E(S) = — XZi21 Pi log2 (pi) (1) 
Where p; is a proportion, can be calculated by 
pi = m which the samples belongs to C; in the total 
sample. (|s| is the total number of sample set S, here 
|s| 2 n). 
The conditional entropy of attribute A: suppose À 
has v different value {a;,a,,..,ay}, the attribute A 
divides the set S into v subsets (S455, ... Sy). nj is 
the sample number of C;. so the conditional entropy 
E(S|A) of the attribute A is: 
E(S|A) = — Xj. pj Xizi pij logz (pij) @) 
Where 
pj is also a proportion, pj — 2 = Bh py is 
a conditional probability, pi; — dsl is the sample 
J 
number that the attribute A belongs to a, in S, 
|S;| = EE, N5). 
The Gain of attribute A: 
Gain(A) = E(A) — E(S|A) (3) 
The GainRatio of attribute A: 
Gain(A) 
GainRatio(A) — Spli(A) 
  
(4) 
Where Splitl(A) 2 — Xj. p; log; (pj). 
C5.0 splits the training samples according to the 
biggest information gain. The first split can define the 
sample subset. Then the second split is according to 
the other field, this procedure will repeat until the 
sample subset can't split At last, check the 
lowest-level split, these sample subsets that has 
non-significant will be eliminated or cut. The key to 
construct decision tree using C5.0 is the training 
samples, choosing a certain number of sample is very 
important. While the number of samples is not the 
more the better, after a lot of experiments we found 
that it is more important for the samples’ Uniformity 
and representative. The other important procedure is 
the feature extraction. The feature mainly include 
spectral and texture feature. The feature’s selection 
should according to the classification system and the 
land cover type. The common feature may contain the 
value of TC, the NDVI, the texture, and so on. 
2.2 Land cover classification workflow based on 
C5.0 
Remote sensing classification depends on the 
theory called statistical pattern recognition, means to 
extract one team statistical feature value of patterns to 
be recognized, and then make the classification 
decision according to one certain rule. The land cover 
classification workflow based on C5.0 has the five 
following procedures: 
1) Establish a classification system 
A suitable classification system is prerequisites for 
a successful classification. Cingolani etal. (2004) 
identified three major problems when medium spatial 
resolution data are used for vegetation classifications: 
defining adequate hierarchical levels for mapping, 
defining discrete land-cover units discernible by 
selected remote-sensing data, and selecting 
representative training sites. In this case, a hierarchical 
classification system is adopted to take different 
conditions into account mainly based on the users’ 
needs. This classification system includes ten ‘level 1°, 
while each ‘level 1’ has some ‘lever 2°. Details could 
be seen in (Higher resolution Global Land Cover 
Mapping Project, 2011). The ten ‘level 1’ includes 
l.artificial, 2.bareland, 3.cropland, 4.forest, 5.grass, 
6.shrub, 7.tundra, 8.water, 9.wetland, 10.Perennial 
snow or ice. 
2) The establish of multiple files 
After remote sensing images were preprocessed 
firstly, and then done the band math, we can get 
feature images, for example, NDVI image, TC image. 
These feature images and the preprocessed images 
were input into the spatial database together, and other 
spatial data can compose one or more multi-band file. 
Selecting what features will depend on the precision 
of result, so the selection of feature images is very 
important. Normally we features present on the image 
have three types of features: 
a. Spectral feature 
Color or grey or the proportion of bands is the 
spectral feature of the target. For example, 
Normalized Difference Vegetation Index (NDVI) is a 
simple graphical indicator that can be used to analyze 
remote sensing measurements, typically but not 
necessarily from a space platform, and assess whether 
the target being observed contains live green 
vegetation or not.