Predicted urban growth map for 2015

6.7 Summary of Results

The NDVI differencing method highlighted a significant decrease in NDVI from 1988 – 2010, particularly in the period 1995 – 2003. After an accuracy assessment was performed it was possible to use the decrease in NDVI as a proxy for urban growth. The general trends identified from the change detection maps are that of a rapidly growing city, with an average rate of growth of 14km² per year over the 22 year study period, but with significant variability – for example the eight years from 1995 -2003 show 163km² of growth, which is over 20km² per year. Over the course of the time-series, urban growth was displayed at increasing distances from the traditional city centre and also close to roads at all times, which displayed a highly significant correlation value. These are two traits very typical of urban sprawl. Using population data it could be seen that sprawl was linked with the growing population in Austin. Although Austin has adopted smart growth initiatives in order to increase its population density it seems that the majority of urban growth is taking place on the urban periphery and Austin appears to be a sprawling city. The decreases in NDVI experienced show that urban growth is taking place at the expense of green/open spaces, which can lead to a fragmentation and deterioration of the natural landscape. Austin is prone to experiencing some of the other issues that surround urban sprawl that have been raised in the past, such as an increase in infrastructure costs at the urban periphery, an increase in car dependency as development takes places further away from the city centre and an increasingly economically stratified city as growth is located in particular areas. Some places have experienced more sprawl than others and the change detection maps and resulting analysis have proved effective in identifying these areas. The resulting maps link well to growth map of the city from the Austin GIS team, which suggests that Landsat imagery and the NDVI differencing method can prove effective in long-term monitoring of urban growth.

The predicted map for 2015 suggests that there will be a total of 431km² of urban growth from 1988 – 2015, with growth of 98km² from 2010 – 2015. The growth appears to be concentrated predominantly in the northern half of Austin. This predicted amount of growth and location does seem excessive, though this has been caused by the model being trained on the changes between epoch 1 and epoch 2, where a large decrease in NDVI was experienced (177km², revised down to 163km² after the accuracy assessment) and the location of this decrease in NDVI was mainly concentrated in the north of the study area. The model for 2010 was validated against the actual land cover map and an over estimation of only 5km² was found, so the change in terms of actual amount was very close. There were lots of misses in location, in particular areas that experienced urban growth in reality that were not picked up by the model. These were investigating further using the false colour composites/aerial imagery and they included a number of areas in West Austin such as Greenshores and Lakeway which are exclusive residential neighbourhoods. As these are not necessarily located particularly close to a primary/secondary road or near an area of previous sprawl it was difficult for the model to predict urban growth in these locations. A time series is presented below with the three change detection maps and the prediction for 2015.

Urban growth: 1988 – 2015 (simulated)

Urban growth: 1988 - 2010

Urban growth: 1988 - 2003

Urban growth: 1988 - 1995

Based on a visual inspection, it would appear that the model for 2015 has over-estimated growth in the northern half of Austin, though in terms of overall amount of change, given the rising population and popularity of Austin this could be a reasonable reflection.

6.8 Limitations

A number of limitations need to be considered in this study. These are presented in the table below.

Table 14: Study Limitations summary

Category	Limitation
Data	Landsat data was used as the core data for the project and the 30m resolution for each pixel means that most urban image pixels will comprise a mix of different surfaces, each having different spectral signatures. This makes classification difficult and it is likely that there would have been spectral confusion in places, which would have impacted on the NDVI results. If available, it would have been improved the accuracy of the study to have used higher resolution satellite imagery. As Landsat data was used, ideally bi-temporal imagery would have been downloaded for each year and a number of researchers, including Wolter et al. (1995), and Yuan et al. (2005) have demonstrated the value of multi-temporal imagery for classification of land cover. Unfortunately due to either cloud cover or quality issues, I was unable to use bi-temporal images, so did ensure that the data selected all had near anniversary dates to decrease seasonal variability.
Methods	The NDVI differencing method used for the change detection analysis presented a challenge in ensuing that change was correctly captured. Changes in NDVI can be directly related to photosynthetic (green) biomass within a pixel, so whilst urban growth will lead to a decline in NDVI, there are a number of other factors which can cause a change in this as well – from a different type of vegetation present to a change in sensor signal at that point. Crucially in this study I was able to filter out agricultural variability. However, it was still difficult to assess real change from noise and set an appropriate threshold. Although the lower threshold worked well and there was little speckle present, areas of increase in NDVI proved more problematic and I was unable to find a threshold to satisfactorily remove these without impacting on true increases. However in terms of capturing urban growth, the NDVI method worked well and proved a time-effective method to capture changes in land cover, without having to resort to a full supervised/unsupervised classification of the data.
Idrisi LCM	The use of the Idrisi Land Change Modeler to create a predicted urban sprawl map for 2015 generated a useful result, however this was a very simplified representation. The input land classification maps were created in the first part of the analysis and subject to issues surrounding classification of pixels, data filtering and sensor changes between change epochs. Only two variables were used in the model to predict change, whilst sprawl is the outcome of a number of intricate, interacting factors. To improve the analysis it would be useful to have had data on other factors, such as land availability, planning restrictions and local economic conditions. A model is a simplified version of reality and can never truly display all of the factors that contribute to urban sprawl, however this exercise does highlight that it is possible to get an idea of possible sprawl amount and direction by using a small number of key explanatory variables. The predicted map was a hard-prediction, which yields only a single outcome of a future scenario, which is chosen from many equally plausible options (Clark Labs, 2009). With this in mind, it can be difficult to achieve an accurate hard prediction for something such as urban sprawl, which can impact upon many different areas.
Reference Data	A final limitation regarding the prediction surrounds the boundary of Austin used in the study. Any future growth was constrained to the limit of the urban area from 2010, when in reality it may be expected that growth would take place beyond this.

Download 0.72 Mb.

Share with your friends: