Remote Sensing and Land Use Classification: Supervised vs. Unsupervised Classification Glen Busch Introduction In this time of large-scale planning and land management on public lands, managers are increasingly looking for faster and less expensive methods of data collection. In efforts to make better decisions, planners need to be able to look at changes over time to assess trends. Policy makers are also looking to assess the effects of policies such as prescribed fire or fire suppression. All of these needs can be filled with data gathered from remote sensors. It is the purpose of this paper to assess the possibilities of using remote sensing for the detection of regional land-use change by developing a land cover classification system. Both supervised classification and unsupervised classification will be tested on a 2000 Landsat image of the spectrally diverse Salt Lake City area. This research will then compare the accuracy of two classification systems at diving the landscape into three classes, water, developed, and undeveloped. Remote Sensing has been used since its inception to group landscape features based on some similar characteristic. Urban areas are no exception. Urban land-use classes determined from remotely sensed data are useful in many applications including land-change detection. However, urban landscapes are heterogeneous by nature with many different land cover types in close proximity. This fact can cause significant error in the classification process resulting in low accuracy among urban sub-classes. Several techniques have been developed to improve the classification process, including expert systems. Theoretical Background Classification Process Before one can begin the classification process, it is necessary to prepare images for the area of study. Care must be taken to properly geo-reference and standardize for the effects of temporal and atmospheric differences between images as well as account for system errors (Ramsey, 2002). Prior to undertaking classification, it is also necessary to define the classes in which to group the landscape. Construct characteristics must be defined in terms of units and scales that the sensor can detect. For instance, the MSS sensor can differentiate between general landscape
classes of forest, grassland, and water on a continental scale. However, if one needs to locate high-density residential areas, it may be necessary to use low-altitude panchromatic images (Jensen, 1996). Jensen also suggests the use of national classification systems, such as the U.S. Geological Survey Land Use/Land Cover Classification System (USGS, 1992) so that results can be compared among similar studies. Having defined the characteristics of the classes to be identified, it becomes necessary to determine the unique spectral signatures of these characteristics that the remote sensor can detect. There are two ways to do this. The first method is called supervised classification and requires the user to locate the desired classes on an image that can be used to train the computer to look for other pixels with similar characteristics. The differences between the spectral characteristics of each class then allows the computer to assign image pixels into the classes that they most likely belong with. The training sets are ideally based on pure signals like that of an alfalfa field or contiguous forest cover. In unsupervised classification, a statistical algorithm separates the image pixels into clusters of similar spectral characteristics. The programmer must then decide how many clusters to have the computer create keeping in mind that more clusters means more spectral differentuation, but also increased effort to assign the clusters to classes. What becomes troublesome is the mixed signals that are given off in transitions between one cover type and another, a phenomenon called sub-pixel mixing (Stefanov et al., 2001). The major loss of classification accuracy occurs when one attempts to assign these mixed pixels into one group versus another. There are other opportunities for error to creep into the classification system. For instance, when there are similar characteristics between classes such as an agricultural field having a similar signature to a golf course and urban grasses. Error can also occur when there is heterogeneity within a class such as the differences between a recently irrigated and dry alfalfa field (Jensen, 1996). To assess the amount of error introduced by mixed pixels and misclassification, it is necessary to compare the results to a separate piece of data, that Jensen calls reference test information (1994). This other information can be an independently created classification, an aerial photograph, or better yet, ground truth. The process essentially works as follows; pixels chosen at random are selected from the classified image and compared to the same location on a reference to determine if the class accurately describes the landscape at the point. This process not only tests the overall accuracy of the classification, it also accounts for the spatial accuracy of the process. The results
of the comparison can then be displayed in an error matrix that quickly shows the site-specific accuracy of the classification process. At this point, the researcher can determine if the results are accurate enough for their purpose, or if it is necessary for more analysis to increase the classification accuracy. Improving Accuracy There are many examples in recent literature of procedures to overcome the misclassification issues. The VIS (vegetation-impervious surface-soil) index developed for urban classification in Salt Lake City (Ridd, 1995), maximum likeliness (Jensen, 1996), and the use of the combination of multiple sensors like that of the ERS-1 SAR and Landsat TM described by Kullich (2000). Expert Systems use a combination of remotely sensed and other sources of geo-referenced data (Stefanov et al. 2001) to increase the information about mixed pixels to make better classification decisions. A study in the greater Phoenix metropolitan area first used a supervised classification of Landsat images in which they replaced the infrared band six with a soil adjusted vegetation index (SAVI). The other data layers (zoning information, water rights, and municipal boundaries) were then used to reclassify pixels of questionable accuracy. For example, a pixel can be reclassified to agriculture instead of a park if it is found to be outside of a city. Using expert systems the study was able to increase average user accuracy from 71% to 80%. Study Area For the purpose of developing a classification system that can be applied to the entire Wasatch Front in later studies, I have subset the Salt Lake Valley from the larger region (see figure 1). The subset includes the land from the top of the Wasatch Mountains in the East to the top of the Ochir Mountain in the West and from Point of the Mountain in the South to Farmington in the North, totaling 813 square miles and providing living space for over one million residents, this diverse landscape contains 11,000 foot mountain peaks, the metropolitan spread of Salt Lake City, and portions of the Great Salt Lake.
Methods The remotely sensed imagery used for the analysis was acquired from a Landsat TM image taken on Path 38 Row 32, September 20, 2000. It was georectified to the WGS 84 spheroid UTM zone 12. The image was then subset to the Salt Lake valley study area. A digital elevation model and the digital-orthophotoquads downloaded from AGRC were also collected at this time for comparisons in later processes. Before attempting a classification it is important to define the categories based on the purpose of the study, here the goal is to divide the landscape into three categories, water, built-up or developed, and undeveloped. Water is apparent, any standing or running water with little or no vegetation is to be put into this class. The developed category is made up of any permanent man-made structure or highly altered landscape feature. This includes both barren ground areas from mining, paved or concrete roads, as well as highly vegetated urban and residential areas. In comparison, undeveloped areas include wetlands, farmlands, parks and forested lands. Figure 1. Greater Wasatch Front Region showing Salt Lake Valley study area Unsupervised Classification To increase the computers ability to differentiate between the three classes one intermediate images was calculated before performing the unclassified classification. Normalized Difference Vegetation Index (NDVI) was first calculated using bands 3 and 4 and was added in place of the missing band 7. This created a 7-layer image that was put though a 30 cluster unsupervised Isodata classification set to.95 convergence level.
With the signature set complete each of the 30 classes were analyzed and placed into the most appropriate of the three categories. Water was the most simple, and the classification process seemed to pick it out well. Putting the rest of the classes into categories was rather difficult as many classes contained pixels of both developed and undeveloped areas Supervised Classification Prior to collecting training sites, seven generalized landscape types were defined; water, irrigated vegetation, upland vegetation, concrete, urban, industrial and fallow/range. After the classification process, the seven classes were combined into the final three categories. For each landscape type I obtained five training sites, which could be averaged into one signature for the type. Using the new signature set (see figure 2 below) the classification was performed. Figure 2. Signature set for seven land use classes, each derived from the average of five training sets. Accuracy Assessment Using the reference test process described by Jensen, a 4-meter resolution digitalorthophotos was used to assess the accuracy of the classifications. The 12 aerial photos were taken in during 1999 and had more than enough detail to compare the 30-meter pixels of Landsat imagery. A random point generator was used to place 50 random points throughout the study area. From here, the classification was compared to the digital image to assess congruency. Results Error Matrix The accuracy of the classification can be described in terms of both producers (omission) and users accuracy (commission). Producers accuracy describes the amount of a landscape category correctly classified on the classification image, while users accuracy describes the probability that
a category on the classification image will be correct when used on the ground. To display both users and producers accuracy, an error matrix is used. The tables below represent the results of the supervised and unsupervised classification process. Unsupervised Classification (30 clusters) Undeveloped Developed Water Row Total Undeveloped 18 9 0 27 Developed 3 15 0 18 Water 1 0 4 5 Column Total 22 24 4 50 Overall Accuracy = 37/50 = 74% Producers Accuracy Users Accuracy Undeveloped = 18/22 = 81% Undeveloped = 18/27 = 67% Developed = 15/24 = 63% Developed = 15/18 = 83% Water = 4/4 = 100% Water = 4/5 = 80% Supervised Classification Undeveloped Developed Water Row Total Undeveloped 10 9 0 19 Developed 0 25 0 25 Water 2 1 3 6 Column Total 12 35 3 50 Overall Accuracy = 38/50 = 76% Producers Accuracy Users Accuracy Undeveloped = 10/12 = 83% Undeveloped = 10/19 = 53% Developed = 25/35 = 71% Developed = 25/25 = 100% Water = 3/3 = 100% Water = 3/6 = 50%
Conclusions Considering the relatively small number of classes used for this research, the results are not exactly impressive. Overall accuracies of 74% and 76% for unsupervised and supervised classification respectively and as low as 50% users accuracy for the water class is less than expected. It was surprised to see that unsupervised classification was as close to supervised as it was because of all the confusion among the 30 sub-groups during the signature process. However, it is here that there is a large potential to increase accuracy in the supervised case. There were at least seven sub-groups that would have benefited from cluster busting. The confusion was especially apparent between the barren lands near the Salt Lake and the high-density urban areas. A major problem with the accuracy assessment was getting the aerial imagery into the some projection as the Landsat image. In many cases, the pixels were counted as misclassified when they were only out of place by one pixel width, likely less than the error introduced by georectification of the aerial photographs. In the future, this is something be aware of. References http://www.gis.usu.edu/~doug/rs5750/lectures/preprocessing_files/frame.htm, accessed 10-10- 02 John R. Jensen, 1996, Introductory image processing: a remote sensing perspective (2 nd ed.), Upper Saddle River, NJ: Prentice Hall, 316p. Kulpich, T.M., and C.C. Freitas, and J.V. Soares, 2000, The Study of ERS-1 and Landsat TM synergism for land use classification, Int. J. Remote Sensing, 21(10); 2101 2111. Stefanov, William L., Micheal S.Ramsey, and Philip R. Christiansen, 2001, Monitoring Urban Land Cover Change: an expert system approach to land classification of semiarid to arid urban centers, Remote Sensing of Environment, v77: 173-185. USGS, 1992, Standards for Digital Line Graphs for Land Use and Land Cover Technical Instructions, Referral STO-1-2. Washington, DC: US Government Printing Office, 60p.