Date of submission: 12
Download 0.72 Mb.
|
- Navigate this page:
- 5.3 Change Detection
5.2 Image Pre-processingGeometric registration is crucial for producing spatially correct change mapping through time. The Landsat data downloaded from the USGS Earth Explorer site had already been accurately rectified and geo-referenced to the UTM projection zone 14N, WGS84 datum. The supporting vector data was also registered to this projection in ArcGIS10 using the Project tool. Once all of the data was in the same projection it was loaded into Idrisi Taiga. Although all seven bands of the Landsat imagery were downloaded from the USGS Earth Explorer website, just bands 1 – 4 of the Landsat data were used for this study. A summary of these bands can be found in the table below: Table 5: Landsat bands summary information
Once loaded into Idrisi Taiga the Landsat imagery was converted from a Geotiff to the Idrisi raster format. An area of interest boundary was generated using the Austin urban area shapefile, which had been clipped from the US urban areas national file in ArcGIS and then loaded into Idrisi. 
Figure 7: Austin urban area, as delineated by the US Census bureau, 2010 The Landsat imagery was clipped to the area of interest boundary using the WINDOW module. A series of false colour composite images were produced for the four years. In displaying a colour composite image, three primary colours (red, green and blue) are used. When the three colours are combined in various proportions, they produce different colours in the visible spectrum. Associating each spectral band to a separate primary colour results in a colour composite image (CRISP 2001). Bands 4, 3 and 2 were assigned to the red, green and blue colour guns respectively. In this image, vegetation appears in shades of red whilst urban areas appear blue. These provide an excellent contrast between natural and man-made areas and were very useful in distinguishing between land cover types. 
Figure 8: Austin false colour composite for 1988 
Figure 9: Austin false colour composite for 1995 
Figure 10: Austin false colour composite for 2003 
Figure 11: Austin false colour composite for 2010 5.3 Change DetectionAs a map highlighting urban expansion had already been created by the City of Austin GIS team (see figure 5) I decided against performing a supervised classification of the data and potentially duplicating this work. Instead, to detect changes in land cover, the NDVI differencing method was employed. Bands 3 (visible red) and 4 (near infrared - NIR) of the Landsat data were used and the NDVI for each image was calculated. The NDVI exploits the circumstance that vegetation has a much higher reflectance in the near-infrared region compared to the visible red region and it is calculated by using a simple red and near-infrared ratio: NDVI = NDVI results are calculated from -1 (low/no vegetation) to 1 (high vegetation). Maps for all of the years were created in Idrisi using this method, by running the VEGINDEX module, specifying bands 3 (red) and 4 (infrared) and selecting the NDVI as the index type.
Agricultural variability has been shown to be an issue in past studies using the NDVI differencing method (Griffith 1998, Masek et al 2000), so the NDVI maps were filtered through the Austin land-use information and agricultural areas were discounted from the analysis. This was achieved in Idrisi by performing an overlay and reclassing the appropriate agricultural areas to zero. This was an important step in ensuring change detected did not include highly variable agricultural areas which would have impacted significantly on the results. Figure 17 below shows the agricultural areas within the Austin urban area boundary. 
Figure 17: Austin agricultural areas, identified from the Austin land use data set After the agricultural areas were removed, the change detection was performed. NDVI images were overlaid and the earlier image NDVI values were subtracted, producing a map of changes in NDVI (ΔNDVI) in which positive values represent ‘greening’ (increased vegetation) and negative values represent ‘browning’ (decreased vegetation). Table 6: Change detection map calculations
A threshold ΔNDVI value needed to be decided upon to successfully capture true changes in NDVI between the two images from noise. It was decided upon a change threshold of two standard deviations from the mean to represent a true change in NDVI value. If the threshold chosen is too small, then any noise such as sensor fluctuation or slight differences in ground reflectance will be viewed as land cover change. Alternatively, if the threshold chosen is too large then true changes may not be identified. Visual checks were undertaken to ensure that two standard deviations did indeed capture true changes and after being satisfied that this was a sensible threshold the NDVI differencing image values were reclassified into three classes: vegetation decrease if a pixel value was lower than the low-end threshold, vegetation increase if the pixel was higher than the high-end threshold and a no change class which contained all values between the thresholds. Full details regarding this method are available in the appendix (see appendix A). As a last step, the change maps were passed through a 3x3 modal filter to remove any isolated pixels, which as Masek et al 2000 note are mostly associated with registration errors. The change detection maps for the three time periods are as presented below. 
Figure 18: Change detection map for epoch 1, between 1988 and 1995 
Figure 19: Change detection map for epoch 2, between 1995 and 2003 
Figure 20: Change detection map for epoch 3, between 2003 and 2010 Download 0.72 Mb. Share with your friends:
|
