RESEARCH ON THE VISUALIZATION SYSTEM AND APPLICATIONS OF
UNCERTAINTY IN REMOTE SENSING DATA CLASSIFICATION
‘National Geomatics Center of China, noi. Baishengcun,Zhizhuyuan, Haidian distric, Beijing
100044, China-keanureevesO 105@ 163 .com
’ Beijing Normal University, Research Center for Remote Sensing and GIS, School of Geography, Beijing Normal
University, State Key Laboratory of Remote Sensing Science, Beijing 100875, China-boyc@bnu.edu.cn
KEY WORDS: Remote Sensing, Uncertainty, Visualization, IDL, Multi-Classifier System
ABSTRACT:
Thematic mapping is one of the major application fields of remote sensing techniques. Data quality of the thematic data was
concerned about. At present, there are a lot of researches that focus on the uncertainty measures, modeling and visualization. But
lacking an easy-to-use visualization system of classification uncertainty. The research work toward how to use uncertainty, especially
how to use it to improve the classification accuracy is few. Based on the review of existing visualization method and system of the
uncertainty of remote sensing data classification, we developed an uncertainty visualization system for remote sensing data
classification. In this system, the static gray, color, color-saturation combined and dynamic 3D visualization method were exploited.
The system was developed using the IDL language. This system can be integrated with ENVI, an popular software for remote
sensing image processing, to use its powerful image processing functions. The system provides users multiple choices of uncertainty
measures, such as: absolute uncertainty, relative uncertainty, relative probability entropy, relative maximum deviation. Its data format
is compatible with ENVI.
1. INTRODUCTION
There are a lot of uncertainties in the thematic classification of
remote sensing. The most common indexes of uncertainty are
confusion matrix and Kappa analysis, which just measure the
uncertainty mathematically but ignore the spatial distribution
characteristics. Therefore it is very difficult to accurately and
truly depict the uncertainty of remote sensing information.
Uncertainty description using visual method is one of the
important part of uncertainty modeling. Visualization
technology allows user to explore the uncertainty of origin data
and relationship among them, which would help them to realize
the effect of uncertainty for project decision.
This paper characterizes the uncertainty information on pixel
scale by various visualization techniques. A visualization
system is developed, which could integrated operate with ENVI.
What is more» besides the traditional visualization techniques
(black and white, gray scale image, colour image et al.), our
system introduced the colour saturation combination techniques
to characterize the uncertainty information, which can let user to
choose colour to represent the class and saturation
corresponding to uncertainty. The results of classification and
distribution of uncertainty would be showed in the
corresponding 2D image.
2. THE ACQUISITION OF REMOTE SENSING
UNCERTAINTY INFORMATION OF EACH PIXEL
The pixel uncertainty information can be obtained from
posterior probability vector during the process of maximum
likelihood classification. Derived from the probability vector of
each pixel, which gained after the classification, can develop
many kinds of measurements of classification uncertainty. The
maximum posterior probability vector can provide a kind of
measurement for classification uncertainty. The higher the
maximum posterior probability, the smaller the classification
uncertainty is. In this paper we introduced four different
uncertainty measurements.
2.1 Absolute Uncertainty
The definition of absolute uncertainty is:
U A (x) = p(o) n \x) /[1 - p{co n |x)
The value range of UA is (0, +oo), the higher this value, the
CO,
In this paper we will introduce some uncertainty measurements,
the visualization techniques, main framework and functions of
system. At last, we will show an example of this system with
Landsat TM data about Lake Lanier in USA.
