Contract Report for HIBECO EU Project
 
 


Download the report (MS-WORD format, 2.9M) and resultant images (Idrisi32 format, 5.6M)

The main task of this contract is to conduct research on image analysis and change detection of satellite imagery for three locations: Vuotso, Utsjoki, and Masi in Northern Scandinavia. In this contract work, a post-classification method is used to detect land cover change from multi-temporal TM imagery. The contract work demonstrates three image processing operations: image preprocessing, false color composite, and multispectral image classification.

The remotely sensed imagery was acquired by TM on Landsat 5 and Landsat 7, covering our research area: Vuotso, Utsjoki, and Masi. Table 1 lists the TM data used in this contract work. The resolutions of original data of Landsat-5 TM and Landsat-7 imagery are 28.5m and 30m respectively. Currently, only Landsat-5 TM imagery acquired on July 24, 1986 and Oct. 08, 1987 covers Vuotso. But there are three dates TM imagery for Masi and Utsjoki. Unfortunately some parts of Landsat-5 TM imagery dated on Aug. 26, 1987 were covered by cloud.

                         Table 1 Landsat data used in the contract work
 
Date
Used bands
Satellite
Covering area
July 24, 1986
1, 2, 3, 4, 5, 7
Landsat-5
Vuotso
Aug. 26, 1987
1, 2, 3, 4, 5, 7
Landsat-5
Masi and Utsjoki
Oct. 08, 1987
1, 2, 3, 4, 5, 7
Landsat-5
Vuotso
June 3, 1995
1, 2, 3, 4, 5, 7
Landsat-5
Masi and Utsjoki
July 29, 2000
1, 2, 3, 4, 5, 7
Landsat-7
Masi (part) and Utsjoki

With constraints such as spatial, spectral, temporal and radiometric resolution, relatively simple remote sensing devices cannot record well the complexity of the Earth’s land and water surfaces. Consequently, error creeps into the data acquisition process and can degrade the quality of the remotely sensed data collected. Therefore, it is necessary to preprocess the remotely sensed data prior to actual analysis. Image restoration involves the correction of distortion, degradation and noise introduced during the imaging process. Radiometric and geometric errors are the most common types of error encountered in remotely sensed imagery. The radiometric and systematic geometric errors of Landsat TM data have been removed by the commercial data provider, while the unsystematic geometric error remains in the image. And also the images are obtained from different dates, therefore the geometric correction is very important. The geometric errors of the Landsat TM data were here corrected by using ground control points before the analysis of land cover change. A ground control point is a point whose position can be determined on the uncorrected image (row and column position) and also on the georeferenced dataset. In this contract work, Landsat-5 imagery head file provides 4 points with georeferenced coordinates which correspond to four corner of the original imagery. The four points have been used as ground control points here. Once the ground control points are collected, the pixels in the uncorrected image are transformed to the georeferenced dataset by means of warping polynomials. Here a cubic polynomial function is used. Each pixel in the corrected image is assigned a new DN value by nearest neighborhood interpolation method. And also the original datasets with 28.5m resolution are resampled into 30m. The resulting average standard errors for the prediction of control points in the master image from those in the slave are less than 0.05 pixel spacing in both row and column (Richards & Jia, 1999). The registered imagery has an exact match with GIS data layers provided by ESRI. As to Landsat-7 TM imagery, there is still a geometric error after geometry correction by using the ground control points provided by the imagery head file. For solving this problem, ten ground control points are selected to register the original Landsat-7 TM imagery to the corrected Landsat-5 TM imagery acquired on 1995. The average standard errors are less than 0.2 pixel spacing in both row and column. In order to perform training data collection, it is necessary to make false colour composite images. The data characteristics for each of the seven bands of the Landsat TM imagery is shown as an example below.
 
TM band
wavelength
characteristics
1
blue
 Designed for water body penetration, making it useful for coastal water mapping. Also useful for soil/vegetation discrimination, forest type mapping.
2
green
 Designed to measure green reflectance peak of vegetation for vegetation discrimination and vigor assessment
3
Red
 Designed to sense in a chlorophyll absorption region aiding in plant species differentiation
4
Near-IR
 Useful for determining vegetation types, vigor, biomass content, for delineating water bodies, and for soil moisture discrimination
5
Mid-IR
 Indicates moisture content of soil and vegetation. Penetrates thin clouds. Good contrast between vegetation types.
6
Thermal IR
 Useful in vegetation stress analysis, soil moisture discrimination, and thermal mapping applications
7
Mid-IR
 Useful for discrimination of mineral and rock types. Also sensitive to vegetation moisture content.

In this contract, a lot of band combinations are tested. We found TM bands 4, 3, and 2 is a suitable combination for collecting training data. It means that TM bands 4, 3, and 2 are combined to make false-color composite images where band 4 represents red, band 3, green, and band 2, blue. This band combination makes birch forest appear as shades of red. Heights will be lighter blue. Water bodies will appear blue. Deep, clear water will be dark blue to black in color, while sediment-laden or shallow waters will appear lighter in color. Urban areas will appear blue-gray in color. Clouds will be bright white. Figures 1a – 1g show TM bands 4, 3, and 2 composite images of three dates (except for Vuotso) for two sites with histogram equalization respectively.
 
 

The maximum likelihood classification (MLC) method was used in this study. The maximum likelihood classification is based on the probability density function associated with a particular training sample statistics. Pixels are assigned to the most likely class based on a comparison of the posterior probability that it belongs to each of the training sites’ statistics. A two-level classification scheme was used in this contract work. The MLC classification was carried out based on the 26 land-use types at the second level of the scheme. Information about the first level was derived by aggregating corresponding types at the second level. Among these land-use types, birch forest types are the most important subjects. The land-use types at first level are listed as following:
1

Water
2
 bogs/wetlands
3
 mixed forests (incl. Pine)
4
 Heaths
5
 birch forest: richer type
6
 birch forest: cranberry-/crowberry-type (with lichencover)
7
 birch forest: blueberry type
8
 Cultivated areas

Traditional supervised training involves the selection of contiguous pixels or blocks of pixel from representative locations across the image as training samples. Selection of the training samples was aided by use of 1:250,000 Maze vegetation map. The eight sets of training statistics were used with the Maximum Likelihood Classifier (MLC). It must be noted that same training sample pixels are used in three date images for each site. Figure 2a –2f show the first-level classified maps of three dates for two sites (class labels refer to above descriptions).
 
 

Post-classification comparison change detection was selected to perform land cover change detection in this contract work. Post-classification comparison change detection is the most commonly used quantitative method of change detection. It requires rectification and classification of each remotely sensed image. These two maps are then compared on a pixel-by-pixel basis using a change detection matrix. The advantage of this method includes the detailed from–to information that can be extracted and the fact that the classification map for the next base year is already complete (Jensen 1996). However, every error in the individual date classification map will also be presented in the final change detection map (Rutchey and Velcheck 1994). Therefore, it is imperative that the individual classification maps used in the post-classification change detection method be as accurate as possible (Augenstein et al. 1991 ). As we know, the different birch forest types are the most important subjects. This work shows some results of change of the different birch forest types between different dates in two sites: Masi and Vtsjokia (since Vuotso only has one date (1986) TM imagery now, this site’s image classification and change detection will be done after the new dataset for other years come). Figures 3, 4, and 5 show the changes of three birch forest types and other land-use types between 1987 and 1995 in Masi resepectively. Utsjokia’s changes between 1987 and 1995 are shown in figures 6, 7, and 8.
 
 
Figure 3a. other land-use types to birch forest: rich types (1987 - 1995) in Masi Figure 3b. birch forest: rich types to other land-use types (1987 - 1995) in Masi
Figure 4a. other land-use types to birch forest: cranberry (1987 - 1995) in Masi Figure 4b. birch forest: cranberry to pther land-use types (1987 - 1995) in Masi
Figure 5a. other types to birch forest: blueberry (1987 ? 1995) in Masi Figure 5b. birch forest: blueberry to other land-use (1987 ? 1995) in Masi
Figure 6a. other land-use types to birch forest: rich types (1987 ? 2000) in Utsjokia Figure 6b. birch forest: rich types to other land-use types (1987 ? 2000) in Utsjokia
Figure 7a. other types to birch forest: cranberry (1987 ? 2000) in Utsjokia Figure 7b. birch forest: cranberry to other types (1987 ? 2000) in Utsjokia
Figure 8a. other types to birch forest: blueberry (1987 ? 2000) in Utsjokia Figure 7b. birch forest: rich types to other types (1987 ? 2000) in Utsjokia