188
CLASSIFICATION OF LAND TYPES IN MINERAL AREAS BASED ON CART
Wenbo Wu, Yuping Chen, Jiaojiao Meng, Tingjun Kang
School of Geomatics, Liaoning Technical University, Fuxin, Liaoning, China, chneyuping317@163.com
KEY WORDS: Remote sensing images, Knowledge, Knowledge based classification method, Classification and Regression Tree,
Land types of the mineral areas
ABSTRACT:
The accurate classification of land types in mineral areas is very important to develop the mineral resources and monitor the
environment of mineral area. Based on the spectral characters and the spatial knowledge of the ground objects in a certain mineral
area in Shenyang which served as a training area. The spectral characters, texture characters, digital elevation model and slope are
selected and extracted. The defined training sample areas are picked up by stratified random sampling techniques based on
geographical coordinates. Firstly, using Classification and Regression Tree (CART) to discovery classification rules which
integrates spectral, textural and the spatial distribution characters from these samples. Then, the interpretation was performed by a
judgement based on these rules. Finally, the traditional supervised— Maximum Likelihood Classifier (MLC) was performed to
check the classification accuracies. The results have suggested that the accuracy of classification based on the CART is higher and it
can obtain a lot of reasonable rules most quickly and effectively. From the highly accurate classification results with multitemporal
Landsat TM images we can detect the change of water bodies, vegetation, road, mining area, residential area and so on. So, it can
provide useful data for the development and protection of the mineral resources.
1. INTRODUCTION
The excessive development of mineral resources is having
initiated the grave resource environment problem. The precise
classification of mining land is essential to develop the mineral
resources and monitor the environment of mineral area. At
present, a lot of scholars inside and outside have conducted the
research regarding this , there are two methods they always
used: One is to make use of aerial photograph or high
resolution satellite image to establish interpretation markt by
lettering, logical analyzing and field inspecting on the same
scale topographic map, then extract the types of terrain feature
by visual interpretation. The classification accuracy of this
method is high, but it wastes time and energy [1]. The other is
to make use of the image features after transforming to build
classification template for each surface feature, and then use the
maximum likelihood classification to classify. The main
alternation methods include Principal Component
Anlysis(PCA), Tasseled Cap, Ratio Method and so on [2] . The
classification accuracy of this method is not high , and it isn’t
applicable for distributed and scattered mining ares.
With commercial high resolution satellite appearing, we will
also be confronted with especially austere challenge by
handling a great deal of remote sensing data while enjoying the
advance of science and technology[3]. How to extract what we
need information and discover implicit knowledge from a mass
of data is the key question of remote sensing image
interpretation at present. Such, data mining and knowledge
discovery technology which emerged in 1980s have being
introduced to the image processing of remote sensing, and the
decision tree as a type of data mining technology was led into
this field just 10 years ago.The decision tree method has many
merits,such as building-up quickly, high accuracy, can creating
understandable trules, the little amounts of calculation. It’s
introduction can make full use of remote sensing data ,so it has
an advantage in information extraction.The principle of
decision tree is not to try to use one kind of algorithm or a
decision rule to classify many category at a time, but is to carry
out the most effective classification specifically for different
aggregation choosing the different standard or method, so it can
oversimplify with complicated problems and resovle
completely [4].
There are many ways to structure decision tree. The
comparatively mature algorithms include ID3 , C4.5 , C5.0
series brought forward by Quinlan, Classification And
Regression Tree (CART) , SLIQ and CHAID brought forward
by Breiman and Friedman.The various algorithms have
respective advantage and deficiency, and ID3, C4.5, C5.0 series
are much uesed in remote sensing field, but these algorithms all
adopt pruning-before, it needs to adjust the parameter carrying
out trial again and again,so this research adopted CART
pruning-after.
2. IMAGE PREPROCESSING
In the study,we chose TM image gained in 11th, August, 2001
of this area as data source.Chose the 1:50000 scale
topographical map of the area as the reference coordinate, and
then carried out geometric precise correction. The amount of
residual deviation was less than 5m, and the error didn’t exceed
one pixel, this result satisfied the study. After the correction, the
size of picture element was 30m, and then selected a trial area
whose size was 256x256 pixel from it. Finally, digitized the
contour lines on the 1:50» 000 scale topographical map, the
contour inteval was 10m, then made use of these contour lines
to creat DEM of the trial area., and then created slope view and
aspect view based on this DEM. The overlay of TM’s false
color composite imagery (RGB543) and DEM is as following
figure 1.
