Land Use / Land Cover Change

Land Use / Land Cover Change

in the Phoenix-Mesa Metropolitan Area

1984 – 2011

Lori Krider & Melinda Kernik

FR 5262

Objectives

The Phoenix-Mesa Metropolitan Area in Arizona, has been identified by the United States Census Bureau as one of the ten fastest growing metropolitan areas in the United States from 1900 - 2000 (1). Increases in the human population often results in changes in the urban/rural dynamics and an increase in stress on water resources, especially in relatively arid regions. Water is the most important limiting resource for occupants of the western United States and it’s allocation is often reflected in how the land is being used. This project is important because knowledge of how the landscape is changing as a result of population growth affects numerous variables, including how the land is managed and how it’s resources are used and preserved.

We used Landsat imagery to classifying land use and land cover in ERDAS IMAGINE 2011 for the time periods of 1984 and 2011 and used these classifications to assess change over time in the metropolitan area. Remote sensing is the ideal application for analyses such as this to take advantage of the spectral signatures of various features on the landscape. Since this area currently has a high population growth rate we expect that changes on the land will be evident in the recent past (~20 years). We also expect to see an increase in impervious surfaces and a decrease in natural vegetation as suburban areas not only become denser where they currently exist but also spread out further from the highly urbanized downtown areas. Lastly, as the population continues to grow, we expect cultivated areas to spread out further from the metropolitan area, also causing a decreasing the natural vegetation. These two important land use / land cover changes will cause an increased stress on already scarce water resources.

Materials, Tools and Concepts

The study area, roughly 16,000 km squared in size, is located in south-central Arizona and contains Phoenix as well as the nearby cities of Mesa, Tempe, Chandler, Gilbert, Scottsdale, Glendale, Sun City, Peoria, and Avondale.

Landsat TM, 30-meter resolution images were obtained from the USGS Glovis (Global Visual Viewer) website for WRS-2 paths and rows 36/37 and 37/37, latitude/longitude 33.2, -111.0 and 33.2, -112.6, respectively, to cover the entire metropolitan area. Landsat TM sensor imagery were chosen over Landsat +ETM imagery due to the scan lines errors present in the +ETM imagery. Although higher resolution, commercial imagery is more desirable for metropolitan classifications, the constraints of the project required us to obtain free imagery rather quickly. The Landsat TM sensor contains 7 bands as follows: blue-green (band 1), green (band 2), red (band 3), near-infrared (band 4), mid-infrared (bands 5 and 7) and far-infrared (band 6) (2). The images are from June 11 and 18, 1984 and June 13 and 22, 2011 to avoid confounding variables of harvest and vegetation coverage. These years were chosen because they represent the earliest and latest available Landsat TM imagery for the month of June with 0% cloud cover and a quality rating of 9. The images are Landsat 4-5 TM product processed by the Level 1 Product Generation System (LPGS). Both sets of images have been corrected to have radiometric and geometric accuracy using group control points from the GLS 2005 dataset.(3)

The 2006 NLCD was obtained from the Multi-resolution Land Characteristics Consortium (MRLC) website, viewed in ArcGIS 10 and used to gather preliminary land cover / land use information. Google Maps™ and Google Earth™ were used as reference to obtain a higher resolution view of features on the landscape. Images where viewed and classified in ERDAS IMAGINE 2011 with the false color composition band arrangement.

Procedures

Landsat TM, 30-meter resolution images were obtained from the USGS Glovis (Global Visual Viewer) website. The bands of each image file were stacked to produce a composite image. All images were cropped to remove the outer, damaged edges. Images of the same time period were merged into a mosaic and cropped once more to include only the Phoenix Metropolitan Area. The 2006 NLCD was used to gather ideas regarding the general land use and land types in the Phoenix area as well as the approximate number of classes that would be reasonable to produce for this area. An unsupervised classification was performed to visualize which classes are commonly confused with one another using a classification scheme with 5, 6 and 7 classes. ERDAS uses the ISODATA classification technique to calculate the means and spectral distances between natural clusters of pixels based on spectral radiometric properties.

A supervised classification was performed using an Anderson-style Hierarchical Classification Scheme at the first and second levels. A combination of levels was used to best represent the natural divisions that occurred on the landscape. The first division included the following general categories: altered, unaltered, developed and water. The altered category was further divided into the following classes: human-assisted and uncultivated. The human-assisted class consisted of the following based on their relative reflectance in the NIR and IR bands and the human influence required for their maintenance: healthy crop and golf courses (reflected shades of red in the NIR and IR bands) and stressed crop (reflected in the NIR and IR bands but as shades of pinks and pinkish-browns). The uncultivated class required no further division and consisted of rectangular areas of land reflected as shades of tan, green, gray and white. The developed class was further divided into the following classes based on relative radiance as a surrogate for imperviousness as well as the relative amount of vegetation contained on each property: suburban, which consisted of mobile homes, r.v. parks and residential housing properties, often including lawns and trees, and urban/roads, which consisted of commercial and industrial properties without lawns and sparse trees and included airports, shopping centers and car dealerships. The water class was also not further divided and consisted of river channels, lakes, and large golf course water hazards. Most of these features were very dark, often black, but a few small, shallow features were greenish and too small to digitize. The unaltered class was further divided into the following classes based on reflectance in the NIR and IR bands and association to water features: natural, which consisted of scrub/shrub vegetation and upland (areas of denser, higher elevation scrub/shrub vegetation interspersed with cobble and boulders, including both shadowed and illuminated hill slopes) and lowland, which consisted of hydrophilic vegetation growing in riparian areas and reflecting in the NIR/IR portion of the spectrum.

A minimum of 15 training areas were defined for each class. However, an upwards of 45 areas were defined for some classes that were produced by a combination of features or were fraught with errors of commission and omission in the first run of the classification. This was not only done to reduce classification errors due to confusion between classes but it was also done in an attempt to capture the variations within classes that were highly variable spectrally. The classes that included more than 15 training areas were human-assisted, uncultivated, natural and urban/roads.

A thematic change detection analysis was performed between the 1984 and 2011 classifications as well as a more general image difference analysis at 10, 20, and 30 percent thresholds.

A stratified random accuracy assessment was performed for both classified images with 20 points per class. No high resolution imagery or field-based ground truth points were available for the 1984 image so assessment was done by referencing the image used in the classification. High resolution NAIP imagery is available for 2007 but was not used in the accuracy assessment because it would have required obtaining 100’s of images to cover the Phoenix-Mesa Metropolitan Area. The assessment for 2011 was done primarily from the source image with reference to Google Maps™ for objects that were difficult to distinguish. In an attempt to reduce bias, the person who assigned the training areas for one year performed the accuracy assessment for the other year.

Results

Our final thematic change detection suggests a significant amount of natural landscape converted into suburban (10%) and urban/roads (4.75%) (a caveat being that natural was frequently confused as being urban in the classification). The analysis also suggests a significant amount of cropland was converted from agriculture to residential/commercial uses. Also, 27.8% of land which was classified as human-assisted in 1984 was classified as suburban in 2011 with an additional 5.5% being classified as urban/roads. A significant amount (19%) of human-assisted changed to uncultivated, suggesting it is still part of agricultural production and simply reflected slight differences in the time the images were collected as well as crop rotation. An additional 17% changed to natural - the magnitude of this change can probably be attributed to errors in the classification. Uncultivated crop showed similar changes, converting to suburban (16.5%), urban/roads (5.9%), and human-assisted (13.1%). Additionally, 12.7% of suburban changed to urban/roads, suggesting increasing density of infrastructure and quantity of impervious surfaces. A large amount of the urban/roads class converted to suburban (23%) and to natural (28%), according to the classifications. This is a reflection of the spectral similarities between suburban and urban classes and demonstrates errors of commission as natural areas were being confused as urban. Only 61.1% of pixels identified as water in 1984 were still identified as such in 2011. Some of this could be due to changes in water levels between the two years (7.2% changed to natural) and shifts in vegetation along riverbanks (6.7% changed to hydrophylic vegetation). Also only 1350 hectares were identified as water in 1984 so small changes in total area have a significant effect on the statistics (Appendix B).

Given the low accuracy of the supervised classifications (described below), the simpler image difference analysis may be more useful in describing the areas of significant urban/rural change across the image. 331870 pixels (around 29,900 hectares) experienced greater than 20% decrease in reflectivity of the NIR band between 1984 and 2011 suggesting significant conversion of agricultural fields to roads and buildings

The overall accuracy for the supervised classifications was 45% for the 1984 mosaic and 67% for the 2011 mosaic (Appendix A). The discrepancy between the two accuracy assessment underlines the difficulty and philosophically tenuous justification of using the same image as reference data. Even having the person who created training areas for 1984 set reference points for 2011 and vice versa, there is a significant opportunity for bias. The basic method of selection training areas was collaboratively established. Strategies differed in the identification of uncertain pixels, such as mixed pixel at the edge of a cultivated field. It is difficult to say how much this difference reflects the relative success of the classifications or the biased or misidentification. By extension, the accuracy of the thematic change detection is also very poor.

The classification would likely have been more successful if a narrower research area with less variation in “real life” classes was selected. This would have resulted in a smaller amount of surrounding natural areas that were not of interest to water consumption changes in populated areas. This could be done by acquiring the 2011 city limits data for Phoenix and the surrounding cities. Interactive maps are available from http://phoenix.gov/webpmo.html and we could have requested the raw data from city officials.

In light of this, our results would be suitable for only a very general understanding of areas of change within the cities but not for detailed zoning/policy planning. We originally planned to make recommendations to city officials regarding water management strategies but decided that our change detection probably would not provide much support.

Discussion

We originally decided to divide classes into the smallest representative unit in order to prevent the classifier from lumping unrelated classes. However, for classes that are spectrally similar, dividing into smaller units can create confusion for the classifier and produce errors of commission and omission. Errors can also be produced if the spectral signature for a class (or classes) encompasses numerous parts of the electromagnetic spectrum, such as a suburban class containing both bright, impervious areas, as well as vegetated and bare ground areas (the latter being more prominent in arid regions).

We hoped to do a broad scale analysis of changes in urban/rural areas (and the associated water use) and, therefore, believed 30m resolution data would be sufficient. Given the difficulties of classifying such a large area, greater spatial resolution would be needed to improve the assignment of classes. Also, acquiring vector files for city boundaries could reduce the relevant study area provided that enough of the rural/urban transition area was considered within city limits. For the specific objective of determining water usage, purchasing better resolution imagery is probably not worthwhile since there are other methods of measuring changes in water consumption which would provide better results than using land use change as a proxy. If, however, the land classification was to be used for multiple purposes, the cost might be justified.

References