Date of submission: 12
Validation and Accuracy Assessment
Download 0.72 Mb.
|
- Navigate this page:
- 5.5 Spatial Analysis
- 5.6 Land Cover Change Modelling
5.4 Validation and Accuracy AssessmentMap validation exercises were undertaken to assess the accuracy of the NDVI differencing method. This composed of two stages. Stage 1 involved estimating the amount of background speckle that were errors of commission. By using the false colour composites, aerial photography and the Austin parks layers I was able to identify areas that have remained as green spaces for the period from 1988 - 2010. Within these regions all decrease in NDVI pixels present could be shown to represent errors of commission, as there should be no change. Figure 21 below illustrates this method of checking for speckle: 
Figure 21: Accuracy assessment for errors of commission – snapshot of epoch 2. The blue circles highlight areas in which a decline in NDVI was experienced, but the area remained as a green space, so these could be seen to be errors of commission. The areas for parks were converted to a raster and reclassed to have a value of 1. The urban growth for the epoch was also given a value of 1 and all other areas had a value of 0. The raster calculator was then utilised to multiply these together, with the areas that returned a value of 1 showing an error of commission. These areas were summed to give an estimate of the frequency of speckle within the growth map for each change epoch. 
Figure 22: Raster calculator to assess errors of commission. The second accuracy assessment exercise was with relation to areas experiencing an increase in NDVI values. This was completed by visual assessment of the false colour composites, aerial imagery and land use information, which contained a category of ‘undeveloped’ land. Similarly to the parks layer, this was overlaid to assess the extent of increases in NDVI against this. As highlighted below, this was a factor and it seemed the spectral signal for undeveloped land would change between the years which had a bearing on the NDVI results obtained. 
Figure 23: Accuracy assessment for increase in NDVI – snapshot of epoch 2. The blue circles highlight areas in which an increase in NDVI was experienced in ‘undeveloped’ land. The red circles are places where NDVI increased but the land use is not undeveloped. 5.5 Spatial Analysis5.5.1 – Location and trends of urban sprawl After completing the validation exercises the general trends and dynamics of urban growth were analysed. A feature of urban sprawl is growth outwards, away from the traditional city centre. Using a Euclidean distance calculation in ArcGIS 10 the distance from the Austin CBD was calculated. The growth from the three epoch maps were overlaid onto this Euclidean distance surface and each pixel displaying growth had a value of 1, which was multiplied by the distance value in the raster calculator. Using this method, it was possible to identify the minimum, maximum and mean distance of urban growth for each epoch from the CBD. 
Figure 24: Calculating the distance of urban growth away from the CBD Another feature of urban sprawl is growth along highways. Using the same method as above, distances of urban growth for each epoch from roads was calculated. As growth away from the city centre and nearby roads are two features highly associated with urban sprawl these were plotted graphically in order to explore them further. A correlation test was undertaken for distance to roads using the Pearson’s product moment correlation and a Student’s t-test performed to test the significance of any relationships found. The census tracts for Austin were also analysed to identify general areas in which the greatest urban change has taken place. The full breakdown of all 304 census tracts is available in the appendix. The 10 census tracts which experienced the most urban growth (as a percentage of their total size) for each change epoch were analysed to identify any spatial patterns, which are presented in the results section. 5.5.2 – Link to the population As a final stage in the change detection analysis the population of Austin over the period of the study was considered in order to assess how the patterns observed in the change detection maps linked with the population in the city. 5.6 Land Cover Change ModellingIn the second part of this study, the NDVI differencing maps were used as an input into the Idrisi Land Change Modeler to produce a predicted NDVI map for 2015. The first step however, involved using the first two change in NDVI maps to create a prediction for the third epoch, as this could then be validated against the actual land cover map for that time to assess the model’s performance. Selection of explanatory variables
Validation of map with actual epoch 1-3 change map  Figure 25: Generalised workflow for the future urban growth prediction map
As the emphasis for the study was on urban growth, just the areas of ‘NDVI decrease’ and ‘No Change (Austin)’ were used as inputs into the model to reduce the number of outcomes possible. Since there were only two land cover classes there were four possible outcomes for each pixel
The change results for epoch 1 (1988-1995) and epoch 1 - epoch 2 (1988-2003) were input into the change analysis tab to highlight the change in the NDVI between the two epochs.  
Figure 26: Change in urban growth between epoch 1 and epoch 2 Between the two epochs a decrease in NDVI was witnessed over a 177km2 area. Download 0.72 Mb. Share with your friends:
|