International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXX V, Part B2. Istanbul 2004 
of China. The landscape in Donghetan is typical in Tarim River 
Valley, with a generally dry and harsh environment, 
represented by typical desert vegetation and soils. With the 
increasing. land development in recent decades, the fragile 
environment has experienced quite remarkable change, largely 
reflecting the general development trend and temporal effect of 
government policies and administrative measures. 
  
  
  
  
Figure 1. Location map of the study area. 
Five multi-temporal remotely sensed images were acquired for 
change detection of this study (table 1), including Landsat MSS, 
TM, ETM and SPOT HRV multispectral images. In addition, a 
multispectral 4-m resolution IKONOS image was acquired in 
September 2000 for field investigation and accuracy assessment 
of image classification. The images were geometrically rectified 
and registered on the map coordinates (table 2). 
Table 1. Data used in this research. 
  
  
  
Satellite Sensor  Path/Row Resolution — Acquisition 
(m) Date 
Landsat 1 MSS 154/31 $7* 3/7/1973 
Landsat 2 MSS 154/31 ST* 12/10/1976 
SPOT 1 HRV  216/266/9 20 20/7/1986 
Landsat 5 TM 143/31 30 25/9/1994 
Landsat 7 ETM 143/31 30 17/9/2000 
  
  
* Resampled resolution. 
Table 2. RMS errors on geometric correction and registration 
of the images. 
  
  
RMSEX RMSEX RMSEY RMSEY 
(pixels) (m) (pixels) (m) 
MSS (1973) 0.23 13.11 0:35 19.95 
MSS (1976) 0.38 21.66 0.49 27.93 
SPOT (1986) 0.21 4.20 0.22 4.40 
TM (1994) 0.24 7.20 0.20 6.00 
ETM (2000) 0.17 4.85 0.16 4.56 
  
2.2 Classification and accuracy assessment 
Using the unified land cover classification scheme developed in 
a previous study (Zhou er al 2004), the multitemporal images 
were classified into five classes including ‘grass and woodland’, 
‘salty grass’, ‘water body’, ‘bare ground’ and ‘cropland’. The 
classification accuracy was assessed using the common 
698 
‘confusion matrix” method, showing an overall accuracy of 85- 
90% with a Kappa coefficient of 0.66-0.78. The details were 
reported by Zhou er al 2004. 
2.3 Change detection 
2.3.1 Measuring the area extent of the change: The 
five-date classified images were integrated to GIS 
database. The area statistics of land use classes were 
obtained from attribute tables. 
2.3.2 Establishing landuse change trajectories: 
Based on the classification scheme, all possible landuse 
change trajectories are shown in figure 2. Note that there 
was no cropland found in this area before 1990's so that 
the class "C" is not included in the classification of 1973, 
1976 and 1986 images. As highlighted in figure 2, for 
example, a trajectory can be specified as G — W — G — 
G — C, meaning that the land was found as 
grass/woodland in 1973, water body (flooded) in 1976, 
grass/woodland again in 1986 and 1994, and cultivated as 
cropland in 2000. 
1973 1976 1986 1994 2000 
  
  
  
  
G s B 
  
  
  
  
  
  
  
  
  
  
  
Cropland Grass and woodland Salty grass Water body Bare ground 
Figure 2. All possible landuse change trajectory identified for 
the study area. 
For the analysis of temporal human impact on the environment, 
we have classified all found trajectories into three generic 
classes, namely, unchanged, stable change and unstable changes 
(table 3). The unchanged class includes trajectories such as G 
— GG G 9 Gand W —^ W —^W 5 W — W indicating 
that the same land cover type was found on the sample point 
over the past 30 years. The stable change class includes 
decisive changes due to human activities such as building 
dam/reservoir and cultivation. They represent the major human 
impact on the environment. The representative trajectories of 
this class include, e.g, G—5 G—5 G—5 C—5C,S5OS2 Go 
G — C, and G — G —5 W — W — W. The unstable change 
class includes those indecisive changes due to the natural 
processes or minor human activities such as light grazing. For 
example, grassland may be flooded during summer and 
subsequently dried out as salty grass because of strong 
evapotranspiration. Examples of trajectories of this class are G 
—> W— B > G — G (flooded, eroded and recovered) and G > 
W—G-— W  G (repeatedly flooded). 
  
The accuracy of the trajectories was assessed using the 
percentage of the ‘true’ landuse trajectories. If at a sample point, 
Intern 
the lai 
referei 
it is a 
schem 
traject 
gener: 
Level 
class 
Unch: 
  
Stable 
  
Unsta 
  
Spatia 
traject 
ground 
Gener: 
trend | 
fluctuz 
the tim 
[If 
bo 
=