• Home
• Research
  • Archaeology/Historic Preservation
  • Digital Library and Data Archives
  • Geospatial Modeling and Visualization (GMV)
  • Environmental Studies
  • Geoinformatics (GIS/RS)
  • Geophysics
  • Spatial / Temporal Social Analysis
  • Urban Infrastructure Projects
  • Visualization/Animation
  • Publications
• Outreach
  • Communities
  • Government
  • International
  • Private Sector
  • University
• Education
  • Digital Institute for Archaeology
  • Field Programs
  • University
  • Continuing and Professional
  • K-12
  • Class Support and Information
  • How do I...
• About CAST
  • Contact
  • CAST 20 Years
  • Our Mission
  • Facilities
  • Computing & Research Equipment
  • Center of Excellence Sponsors
  • Collaborators
  • Staff Web Pages
  • Publications
  • CAST in the News
  • Our Community
  • Directions to CAST

• A - Z
• Support

Center for Advanced Spatial Technologies

University of Arkansas

Change Detection

Home » Research » Environmental Studies » LULC Change in Carroll County » Methodology » Change Detection

Chapter 4 - Methodology

Data Acquisition  |  GPS Ground Truth Data Acquisition  |  Image Preprocessing  |  Image Classification  |  Change Detection  |  1984 Land Use & Land Cover Map  |  Data Conversion  |  Land Assessment  |  Hardcopy Map Production

Change Detection Technique

Image differencing is probably the most widely applied change detection algorithm (Singh, 1989). It involves subtracting one date of imagery from a second date that has been precisely registered to the first. According to recent research, image differencing appears to perform generally better than other methods of change detection (Coppin & Bauer, 1996). A primary goal of this research was to utilize a current TM-based land use and land cover image from which to detect prior forest conversion to pasture lands by employing an image differencing technique. The change detection technique employed in this study is based on the previous research by Maus et al (1992), Doak and Lackey (1993), and Green et al (1994).

The 1984 land use and land cover map could not be classified in the same manner as the 1999 map due to the lack of GPS ground truth data for training site development and accuracy assessment; so, a land use and land cover change map was first produced by creating a difference/change image file employing a Band 7 subtraction of the 1984 TM image from the 1999 TM image. (Figure 4.9)

Figure 4.9

Band 7 has been shown to yield successful results in identifying change in land cover in forested areas (Maus et al, 1992). This portion of the electromagnetic spectrum (middle-infrared, 2.08-2.35 um) is less sensitive to atmospheric variation helping to minimize undesirable radiometric variation (Doak & Lackey, 1993) between the 1984 and 1999 Landsat TM images. By employing only 2 bands in the change detectionprocess, data redundancy issues, prevalent among vegetation indices, and the addition of potentially misclassified change errors were minimized in the difference file. In addition, previous research has shown that a Band 7 subtraction file to be more consistent at identifying vegetation loss as compared with TM Band 4/3 vegetation index (Green et al, 1994).

Prior to subtraction, both bands were evaluated for minor differences in reflectance unrelated to changes in cover type. This was accomplished by evaluating histograms for each Band 7 loaded in different image planes. The 1999 Band 7 had a mean value of 25.7 while the 1984 Band 7 had a mean value of 18.4, a difference in value of 6.3. This number was rounded up to 7.0 in order to ensure optimum normalization between both Band 7’s. This process aided in the reduction of in-between scene variability as a result of potential differences in atmospheric conditions during satellite scene acquisition. Figure 4.10 shows the histograms of Band 7’s from both the 1999 and 1984 TM imagery.

Figure 4.10

Values in the 1999 Band 7 image were therefore shifted downward by a value of 7 before subtraction and a constant value of 100 was added to the difference file values to compensate for software limitations prohibiting negative values. The band 7 difference file then was created as:

Δ_ij = ( V7_99ij - 7 ) - V7_84ij + 100

Where

Δ_ij = Change in Pixel Value

V7_99ij = Reflectance Value of TM Band 7 in 1999

V7_84ij = Reflectance Value of TM Band 7 in 1984

i = Row Number

j = Column Number

The difference file was output to a 16 bit image channel in order to maximize the range of potential change values. An 8 bit image channel is limited to 256 values ranging from 0 – 255, where as a 16 bit image channel has a range of 65,536 values. The maximum change value from the difference file was 297 in the 16 bit image channel. The difference file was characterized by a mean value of 105.0 and a standard deviation of 13.6.

The difference file was visually analyzed to determine relationships between difference file values and changes in vegetative cover between dates. Upon viewing both 1984 and 1999 images displayed in band combinations 7, 5, 3 (RGB) and 7, 4, 3 (RGB) at full resolution, the Band 7 difference file demonstrated to be an excellent indicator of level I, forest clearcuts and subsequent conversion to pastures. The detection of increases in vegetative cover associated with the regeneration of forests from fallow pasture was also visually discernable. This finding was an unexpected but pleasant discovery, although should be a logical deduction in the grand scheme of land use change.

Other land use changes were also further investigated such as urban growth. However, the assessment of land cover change in non-forested areas is often more difficult due to both the rapid occurrence of change and the less-definable relationships between spectral change and land cover change (Green et al, 1994). The strength of the Band 7 subtraction was limited to detecting changes in brightness values as a result of significant changes in vegetative cover. Other land use and land cover change potentially detectible from this technique were beyond the scope of this study. The subtraction of Band 7’s of the two images capitalized on the ability of that band to overcome potential atmospheric interference and to readily identify vegetation loss.

A critical element of the image differencing method is deciding where to place the threshold boundaries between change and no-change pixels as displayed in a histogram (Singh, 1989). Singh (1984) also suggests selecting thresholds on the basis of the number of standard deviations from the mean value of the change class under investigation.

The determination of change threshold values was determined by the visual comparison of the difference file with both the 1984 and 1999 TM scenes and subjective manipulation of change values to represent both positive and negative changes in vegetative cover based upon the mean and standard deviation of the difference file. Preliminary analysis was conducted utilizing +/- 1 standard deviation from the difference file mean. However, this did not adequately represent all potential forest change areas displayed in the 16 bit image channel. Upon further analysis and visual interpretation, it was decided that +/- 2 standard deviations (27.2) from the difference file mean (105.0) effectively portrayed all areas of potential change. These values were calculated and referenced to the 8 bit range of values (0 – 255) for further analysis. The numeric range of the change categories were labeled as follows:

     1 – 83 = Vegetation (Forest) Gain

     84 – 131 = No Change

     132 – 255 = Vegetation (Forest) Loss

The division of these change categories is relative when considering the reflectance properties of the Earth’s surface as captured by the Landsat TM sensor. Areas classified as vegetation gain (forest regeneration) are generally characterized by lower reflectance values (Digital Numbers) as compared to areas classified as vegetation loss (forest clear cuts) are generally representative of brighter reflectance values.

There was no measurable quantitative amount of vegetation loss or gain, in a percentage, associated with each of the 3 vegetation change categories. Areas of vegetation gain were characterized as forest regeneration from fallow pastures, and areas of vegetation loss were characterized as forest to pasture conversion from clear cutting.

In order to determine the extent of forest and pasture conversion that had transpired over the temporal span of this study, the Band 7 difference file was classified using an unsupervised, ISOCLUS, classification. This classification incorporated Bands 2, 3, 4, 5, and 7 from both the 1984 and 1999 images in order to optimize classification results. The areas that had been designated as experiencing change, were classified into 224 different cluster categories which were subsequently aggregated into 3 classes: forest to pasture conversion, pasture to forest regeneration, and no change. All other potential changes detected by the Band 7 difference file were classified as no change to coincide with the scope of this study.

The final Band 7 difference file was filtered by both a 5 by 5 mode and 5 by 5 sieve filters in order to remove all areas of change less than the minimum mapping unit of 1 acre (approximately 5 pixels). The resulting land use and land cover change map, 1984 – 1999, adequately portrayed both the conversion of forest to pasture from clear cutting and the regeneration of forest from fallow pasture.