Study Area and Data
The study area is located in the west of Yuli County, Xinjiang Uygur Autonomous Region of China. The 64 km2 study area is centered at about 41°5′N and 85°43′E and located in the middle reach of the Tarim River, the longest inland river of China (Fig. ). At the fringe of the Taklimakan Desert, the “green corridor” of the Tarim Basin is one of the most important habitation areas in the arid zone of China. The landscape is generally characterized as a dry and harsh environment, represented by typical desert vegetation and soils. The area is dominated by a floodplain with elevations ranging from 900 to 910 m.
Location of the study area
Natural vegetation consists of trees, shrubs, and herbage. Tree species includes Populus euphratica, shrubs include Tamarix spp., Nitraria sibirica, and Halimodendron halodendron (Tamarix ramosissima, Tamarix hispida, and Tamarix elongata are dominant species), and herbage includes Phragmites communis, Poacynum hendersonii, Alhagi sparisifolia, Glyzyrrhiza inflata, and Karelinia caspica, etc. The main crop is cotton, which is planted in April and harvested at the end of September. With increasing land development in recent decades, the fragile environment has experienced quite remarkable change, largely reflecting the general development trend and temporal effects of government policies and administrative measures.
Five multi-temporal remotely sensed images were acquired for change detection in this study, namely, Landsat MSS (3/7/1973, 12/10/1976), TM (25/9/1994), ETM+ (17/9/2000), and SPOT HRV (20/7/1986) multispectral images. All five images were used for the analysis of categorical changes to establish the categorical change trajectories of land cover types, but only three of them (1973, 1986, and 2000) were used for the quantitative analysis of vegetation condition change to avoid potential systematic biases in computing vegetation indices. In addition, a multispectral 4 m resolution IKONOS image acquired in September 2000, which covered the whole study area, was also used to assist field investigation and accuracy assessment of image classification. The IKONOS data were registered to the geo-referenced aerial photo on the scale of 1:10,000. This geo-coded IKONOS image was then taken as the master and the other remote sensing data were rectified by image to image, with an average registration error of less than half a pixel. To make the classified land cover images comparable, the Landsat images were resampled to 20 m, which is the resolution common to all of the images, using the nearest neighbor method.
A reference data set of 790 sample points was obtained. We chose a stratified random sampling scheme based on a 2000 ETM+ classified image, which had the most detailed land cover pattern. We collected the reference data of 2000 ETM+ directly based on the 2000 IKONOS image and fieldwork. For the images acquired on the other four dates, we used the IKONOS image as the basis to locate sample points. Comparison was then made with satellite images acquired on other dates for each of the sample points. By this means, obvious land cover changes such as grasslands to water and bare ground to cropland could be reliably detected by image interpretation. Field visits and interviews of elderly locals were also conducted for sample points where a clear relationship between the present and historical images could not be established.
Land Use Change
In arid zones, land use change can be characterized as two kinds: categorical or quantitative changes, which can also be further classified as reversible or irreversible changes. Irreversible change means that land use features have changed into other types and generally cannot change back (e.g., dam construction causing grassland to be permanently flooded). Reversible change, on the other hand, means that the changes have not reached the “no return” stage so that the original status of the land cover can be restored (e.g., vegetative cover change due to weather conditions or seasonal flooding).
For the analysis of categorical change, the approach of change trajectory analysis (Miller et al. 1998
; Larsson 2002
; Yang and Lo 2002
, Zhang et al. 2002
; Zhou et al. 2008a
; Li and Zhou 2009a
) was used. Supervised classification was employed to classify individual images independently, using a unified land cover classification scheme to ensure that the classifications of the multi-scale, multi-temporal images are compatible with each other. The land cover was grouped into five categories including cropland (mostly irrigated cotton fields with a full green vegetation cover in the summer), grass and woodland (mostly native pastures with a sparse vegetation cover typically less than 50%), salty grass (salt-tolerant vegetation types at the fringe of water bodies or abandoned croplands), bare ground (e.g. sand dunes), and water bodies. The classified images were then combined in a GIS to establish the categorical change trajectories, e.g., grass and woodland (1973), grass and woodland (1976), grass and woodland (1986), cropland (1995), and cropland (2000). The identified land use trajectories were then grouped into three generic categories, namely, unchanged, human-induced, and naturally changed (Zhou et al. 2008a
). Under this classification system, we considered the human-induced change as irreversible. In contrast, the natural changes were considered reversible, i.e., the original status of the land cover can be restored when conditions permit.
The unchanged class indicated that the same land cover type was found on the sample point over the past 30 years. The human-induced change class included decisive changes due to human activities such as the building of a dam/reservoir and cultivation. Old cultivation indicated that land cover had changed to cropland prior to 1994 and has since remained as cropland. New cultivation indicated that land cover changed to cropland at some time between 1994 and 2000, and in 2000 remained as cropland. Reservoirs/ponds indicated that land cover changed to and remained as water bodies since 1986. These changes were often irreversible so that they represent the major human impact on the environment. The natural change class included 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 grassland because of strong evapotranspiration. Grass/woodland indicated that land cover changed periodically between grass/woodland and salty grassland. The flooded category indicated that land cover had changed periodically between water and other land cover types. Bare ground indicated that land cover changed periodically between bare ground and other land cover types.
Quantitative change evaluated the conditions of vegetation that resulted from temporary natural factors and that allowed the original status to be restored. Naturally, the irreversible changes (i.e., the human-induced change category as specified above) were excluded from this quantitative change analysis. The Normalized Difference Vegetation Index (NDVI) was used to compare and analyze the quantitative change of vegetation. NDVI is sensitive to the presence, density, and condition of vegetation and was correlated with absorbed Photosynthetically Active Radiation (PAR) and vegetation primary production (Herrmann et al. 2005
). In spite of the impact of the vegetation phenology, the moisture conditions, the sun zenith angle or sensor view angle, and the differing wavelengths of different sensors, NDVI was well suited to the study of vegetation greenness in arid zones (Olsson et al. 2005
In this study, to avoid uncontrollable systematic bias in computing NDVI, only three of the five available images (1973, 1986, and 2000) were used because they were all acquired in the summer season without substantial temporary effects (such as flooding). The 1976 image was excluded because of its late acquisition date (in the autumn when the vegetation phenology significantly varied from that in the summer) and the 1994 image was not used due to the extensive cover of flood water.
- Normalization of Remote Sensing Data
In order to make a quantitative comparison between digital images, radiometric normalization was undertaken to eliminate the radiometric and atmospheric effects on the images. Two approaches to radiometric correction are possible, namely, absolute and relative methods. The absolute approach requires the use of ground measurements at the time of data acquisition for atmospheric correction and sensor calibration. This is not only costly but also impractical when archival satellite image data are used for change analysis (Hall et al. 1991
). The relative approach (Yang and Lo 2000
), which does not require simultaneous ground data acquisition, is therefore, preferred. Various methods are available for the relative approach to radiometric normalization (RRN), such as robust regression (Olsson 1993
) or the use of invariant target sets (Eckhardt et al. 1990
; Jensen et al. 1995
; Michener and Houhoulis 1997
), pseudo-invariant features (Schott et al. 1988
; Henebry and Su 1993
; Yang and Lo 2000
), or a Radiometric Control Set (RCS) (Hall et al. 1991
In this study, the RCS method was used. Two sets of extreme features with nearly invariant reflectivity and identifiable on both image scenes were collected as targets, which should be independent from seasonal or biological cycles (Schott et al. 1988
; Hall et al. 1991
). The water body and bare ground were selected to represent the dark and bright control sets, respectively. The mean of the pixel digital numbers (DNs) was then derived for each of the normalization spectral bands to derive parameters used in linear regression as:
was the DN of band k
in image X
on subject image,
is the DN of band k
on normalized image, mk
is the slope and bk
is the intercept. The parameters mk
could be calculated as:
was the mean DN for the bright control set of the reference images, Bsk
was the mean DN for the bright control set of the subject image, Drk
was the mean DN for the dark control set of the reference image, Dsk
was the mean DN for the dark control set of the subject image, and k
is the band number.
The control sets were determined by intersecting post classification images instead of the Kauth–Thomas (KT) scattergram isolation method that was used by Hall et al. (1991
). All five epoch images were classified using ISODATA with 15 classes and 500 iterations. After labeling the bare ground and deep water and intersecting the five epoch images, the common area of target samples of water and bare sand were 300 and 5,452 pixels, respectively. The SPOT HRV image of 1986, which was the median of the time-series of the multi-temporal images, was used as the reference data. The normalization parameters and adjustment on derived vegetation indices were as shown in Table .
- Vegetation Coverage Change
The normalization parameters and NDVI adjustment
Vegetation Index (VI) is commonly used for evaluating vegetation condition. In this study, we employed the Normalized Difference Vegetation Index (NDVI) for this:
Because of the availability of remote sensing data, the 1973 Landsat MSS, 1986 SPOT HRV, and 2000 Landsat ETM+ images were used to analyze vegetation conditions over the study period. The NDVI was calculated using Bands 5 and 7 of Landsat MSS, Bands 2 and 3 of SPOT HRV, and Bands 3 and 4 of Landsat ETM+ as red and near-infrared bands, respectively. The resulting NDVI images were then compared and thresholds were used to categorize the vegetation change.
In this study, the plant condition changes were taken into account. The high NDVI means better vegetation conditions. Where NDVI is less than or equal to zero (NDVI ≤ 0), the land cover types are most likely bare ground without vegetation such as sand dunes and water bodies, so that they are excluded from the quantitative analysis. Moreover, the area of human-induced categorical change was also excluded.
Climatic, land use, and socio-economic data were collected for the investigation of the effects of natural and human factors on land cover change. In the arid zone, precipitation was the most critical natural limitation, while human factors included direct impacts such as the construction of irrigation works and agricultural practices, and intrinsic impacts such as population growth and change of land use policy. Annual precipitation data in Korla City, which is close to the study area, were acquired from the China National Meteorology Bureau. The irrigation works such as reservoirs, dams, ditches, and dykes were mapped by interpreting high-resolution IKONOS imagery, and their construction dates were acquired by field investigation. The pattern of agricultural practices was investigated by comparing the spatial pattern of farmland over the study period and by field survey. The change of local land use policy was studied by interviewing local elderly farmers and governors. The correlation between the human activities and land cover change was then analyzed.