Haidas, 
«»text pe r . 
seen for all 
led w ith the 
Ending pi x . 
loca l condì- 
ma ge resolu- 
lth m was in 
§ er than for 
'therworkis 
»crease ela,. 
“tihcation. 
n. Ed. P. 
»servation ( 
istics. 1EF 
A Method for Proportion Estimation of Mixed Pixel(MIXED 
by Means of Inversion Problem Solving 
Kohei Arai and Yasunori Terayama 
Information Science Department 
Science and Engineering Faculty 
Saga University 
1 Honjo, Saga-city, Saga 840 Japan 
ABSTRACT 
A basic idea of a contextual classification by means of a 
method for proportion estimation of Mixed Pixel(MIXED featuring 
Inversion Problem Solving technique together with a concept of a 
Image Classification Artificial Intel 1igence(ICAI) with a 
hierarchical multi-information integration are proposed. 
An example with Landsat-5 TM data shows that estimated 
proportions are useful as a contextual information when the pixel 
of interest is MIXEL and is classified to an undesirable class in 
the spectral and spatial feature space. 
Introduction 
2. Basic concept of 1CA1 system with a hierarchical 
multi-information integration 
As well known, not only spectral and spatial 
information derived from remote sensing satellite 
imagery data but also relational information such as 
inclusion, connection, etc. can be used in image 
classification method. Fig.1 shows a hierarchical 
structure of multi-information to be used for image 
classification. In human perception process, three 
layered multi-information are referred in almost 
same time. The proposed 1CAI system should include 
such information with flexible accessibility and 
with traveling among the layers back and forth while 
backward and forward reasoning should be capable. 
Discontinuities of roads are sometime observed 
in remote sensing satellite data. It depends on the 
relationship between Instantaneous Field of 
View(IFOV) and road width and/or the angle between 
road and scanning directions. The pixels in the 
discontinued portion of roads tends to be classified 
as an improper class in such case because of their 
properties in spectral and spatial feature spaces. 
Contextual information is useful in such situation 
together with information on mixing ratio or 
proportion of each class in the pixel of interest. 
Not so many papers dealt with proportion 
estimation were published(Ref. 1 - 3). Namely the 
method based on the feature mixing concept, the 
method with a generalized inversion matrix, the 
method based on the second order programming, the 
least square mean of method and linear regressive 
analysis with surrounding 8 neighbor pixels. The 
proposed method is based on Inversion Problem 
Solving method in a straight forward manner. 
The proposed method is one of the elements of 
the Image Classification Artificial 
Intel 1igence(ICAI) with a hierarchical multi 
information integration proposed herewith. 
First, a basic concept on the ICA1 is described 
followed by proposition of the proportion estimation 
method using Inversion Problem Solving. Then an 
example with Landsat-5 TM data of Saga, Japan is 
shown. 
For instance homogeneous segments may be 
classified using only spectral information. On the 
other hand, spatial information should be considered 
for heterogeneous segments. It can be divided into 
two categories, microscopic and macroscopic 
properties. The former is based on relationships 
between the pixel of interest and its surrounding 
pixels while the latter is a spatial structure in 
the image of interest such as road network, etc. 
Contextual classification and spatial treatments are 
useful for such pixels. Beside these relational 
information can be also used. 
Contradictions are sometime observed between 
classified results obtained through spectral and 
spatial classifications. For instance, the 
aforementioned discontinued segments between 
continuous road segments are classified as the 
classes other than road class because of their 
spectral signature. Sometime roads are covered with 
crown of trees. Beside this, narrow width roads 
compared to IFOV tend to be discontinued in the 
image depending upon the angle between the road and 
scanning direction, ratio of road width to IFOV, 
etc. In such situation, contextual information 
should be taken into account together with 
relational information. The proposed ICAI system 
allows us to refer such information including in the 
hierarchical layered structure.