HIGH ACCURACY AUTOMATIC CLASSIFICATION OF MULTI TEMPORAL DATA 
Haruhisa SHIMODA, Kiyonari FUKUE, Tukasa HASHINO, 
and Toshibumi SAKATA 
Tokai University Research and Information Center 
2-28-4 Tomigaya, Shibuya-ku, Tokyo 151, Japan 
ABSTRACT 
A new automatic landcover classification method for multi temporal data is presented 
in this paper. This method is based on prior condition that a landcover map of the 
target area already exists, or at least one of the multi temporal data had been 
classified. The landcover map or the classified result is used as a reference map to 
specify training areas of classification categories. The new classification method is 
consist of five steps, i.e., extraction of training data using the reference map, 
change detection based upon the homogeneity of training data, clustering of image data 
within areas where changes were detected, reconstruction of training data using 
overall image data, and maximum likelihood classification. 
In order to evaluate the performance of this method, temporal Landsat TM data were 
classified by using this method and a conventional method. As a result, we could get a 
landcover change with high reliability and fast throughput, without a skilled 
operator. 
Key Words; multi temporal data, change detection, landcover classification, 
automatic training data selection 
1 INTRODUCTION 
Since the launch of Landsat-1 on 1972, 
enormous observation data have been 
integrated in various organizations such 
as receiving stations, institutes, and 
application users. Many studies have been 
trying in order to utilize those data 
with more effectively, e.g. EosDIS(Earth 
Observing System Data and Information 
System) and GIS(geographical Information 
System). It is assumed that processing 
technology for multi temporal data have 
become important from a view point of 
above the present condition. 
Utilization objectives of multi temporal 
data can be divided into two categories; 
(1) to improve analyzing accuracies, and 
(2) to analyze change of the earth surface 
such as land-cover/use, and sea surface 
temperature. This research presented in 
this paper is concerned about the later 
objective, especially for land-cover/use 
changes. 
Following two methods have been usually 
used as change detection methods of land- 
cover/use; (l)a method detecting change 
areas from observation image data 
directly, and (2)a method detecting those 
from multi temporal thematic maps. In 
general speaking, the first methods can 
not detect change areas exactly. One of 
the reasons due to that it is difficult 
to compensate influences of variation 
about observation conditions such as sun 
elevation, atmospheric condition, and so 
on. The other one depends on that 
spectral features of a land-cover/use may 
change according as seasonal changes, 
e.g. leaf color of forests, grass 
rankness and water contents of bare 
grounds, growing situation of crops, and 
etc.. On the other hand, the second 
change detection method has potential to 
eliminate those problems. 
However, there are several problems on 
conventional supervised methods of land- 
cover classification for a set of multi 
temporal image data. First, training data 
must be extracted by a skilled operator 
for each temporal image in order to get 
high classification accuracies. Second, 
the operator should perform ground truth 
before training data extraction. As well 
known, ground truth and training data 
extraction are highly time consuming 
processes. Third, it is difficult to 
extract sufficient training data using 
supervised method especially for high 
ground resolution sensor data such as 
Landsat TM and SPOT HRV, because of large 
variation of image level with pixel-by- 
pixel/1,2/. For instance, value of pixels 
in low density urban areas have very 
different values by a reason which those 
area are consisted of many kinds of 
landcovers, e.g. house roofs with several 
colors, lawns, trees, concretes, 
asphalts, bare grounds, etc.. Further, it 
is difficult to analyze landcover changes 
because of unstable classification 
accuracies for each temporal data. A 
method overcoming these problems for 
conventional supervised classifications 
is presented in this paper. 
2 PROPOSED AUTOMATIC CLASSIFICATION METHOD 
The major problems on supervised methods 
of landcover classification for a set of 
multi temporal image data are induced 
from supervised training data extraction, 
as described at previous section. There 
fore, in order to solve those problems, 
it is necessary to automate the process 
of training area selection and adjusting 
of training data according to each multi 
temporal image. 
228