COMPARISON OF LANDUSE CLASSIFICATION SYSTEMS FOR LOGAN

We focused on the southwest quadrant of city for two main reasons. First, this is one of several margins of the city where growth has been most expansive. And second, the contrast between urban and rural land uses is the greatest in this area.

We then looked at Logan on a property parcel level which required the creation of a coverage from the Cache County tax plat maps. The maps were obtained from the County Recorder's Office, and were then digitized into ARC/INFO. The TIC file of an existing coverage was used as a base for our coverage. The four tics from this coverage were in State Plane Coordinates and the entire extent of the City lies within the four coordinates. The City's engineering software package, PALETTE, was used to get real world coordinates on the plat maps. From PALETTE, features that already had State Plane Coordinates, such as sidewalks, fences, and intersections, were used. Common points from PALETTE and the plat maps were ascertained and a FORTRAN program was run on them which took the points and converted them into a .gen file. This file was then FTP'ed to the ARC/INFO workstation where the coordinates were then converted to a TIC file. The digitizer was then coordinated and the plat maps were digitized into separate coverages. At this point, all the coverages were combined into one parcel coverage of our area of interest.

The next step after cleaning our parcel coverage and building topology, was to attribute it with land use information visually interpreted from the aerial photography. A large map of our coverage was plotted so that we could use it to transfer land use information from the photos. The classification system we used was based on the General Land Use Codes which were developed by the American Planning Association. The interpretation of air photos was conducted by one person to maintain a more uniform classification. Once the air photo interpretation was completed, we took the map and entered the land use code into the land use item in the .PAT. Finally, the GIS coverage was reprojected from State Plane Coordinates into UTM, in order to better compare the results of the GIS and RS classifications.

The SPOT scene covered an area larger than our area of interest, so it was necessary to subset out the areas outside the AOI. We concentrated on an area of Logan City which included everything west of 300 East, and south of 200 North.

This original image was not geometrically corrected to a real world coordinate system, so georectification was required before further analysis could be performed. The following chart illustrates what occurs during the rectification process.

One advantage of using a DOQ was the one-meter spatial resolution, which was previously registered to UTM coordinates. This enabled us to easily identify ground control points for georectification of the SPOT subscene. In addition, the DOQ enabled us to obtain a low Root Mean Square error of .03m. The disadvantage of this image to image registration is that our SPOT subscene will contain any registration errors that the DOQ contains. We will assume that these errors are not significant to our study of landuse, and continue on to image enhancement and classification.

An unsupervised classification was attempted but proved useless to our classification study.

A supervised classification was explored next using taining sites and weighted probabilities and the maximum likelihood algorithm. The weighted probabilities were adjusted and readusted in order to provide a reasonable classification. Training sites were selected for five classes, namely, commercial, agriculturre, mixed residential, industrial, and water.

The supervised classification was performed on the SPOT image. The results were then converted to an ARC/INFO coverage. The identity function was then used to combine the image and GIS classifications. The following image indicates where the classifications between the two matched.

The table below is an accuracy assessment on the comparison between the two classifications. Twenty-five points were selected for each class and the image classification was compared against the actual land use at each of these points. (It should be noted that our sample size was quite small for the number of polygons. Some of the polygons may have been very small where land use matched, and some may have been larger where land use mismatched, therefore the total area where land use actually matched may not be represented in the accuracy assessment table.)