An unsupervised classification (MMU .09 ha) of four principle components was also performed using the same 1993 TM scene used for the Williams classification. Only four classes were delineated: Oakbrush, Sagebrush, Juniper, and Other. An overlay of all three datasets for the Oakbrush class revealed a near exact agreement between the principle components and Williams classifications with spotty inclusion errors in the GAP dataset probably due to generalization and a large area of exclusion attributable to a fire occurrence between imagery dates (GAP 1988 and Williams 1993).
The goal of this research is two-fold. The first is simply to identify the the amount of agreement between the two datasets, which is accomplished through both qualitative and quantitative methodologies. The second goal is to identify the impact of the MMU and of separate classification dates. This is accomplished by reclassifying the 1993 imagery at the .09 hectare MMU using a technique independent of either classification.
Coverage Integration:
The next step is to generalize the Williams .09 hectare GRID
coverage into 1 hectare cells. The smart-raster-eliminate
(Bassett, 1995) is used with an ecologically based elimination
matrix:
12 12 7 6 11 9 10 1 8 5 1 -1 1 7 12 8 9 11 3 3 3 10 3 -1 3 11 5 12 5 5 5 5 5 5 5 -1 10 11 9 5 12 11 10 5 5 5 5 -1 5 5 5 5 10 12 11 5 5 9 5 -1 5 5 5 5 5 11 12 5 5 10 5 -1 5 5 5 5 5 5 5 12 11 5 5 -1 5 5 5 5 5 5 10 5 5 12 11 -1 9 5 5 5 5 5 5 5 5 11 12 -1 10 5 5 5 5 5 5 5 5 5 5 12 5 10 5 10 5 5 5 5 10 10 -1 12
After the GRID is generalized to the MMU, it is then polygonized using the GRIDPOLY command in ARC. If the GRID is not in integer format, it first needs to be converted inside GRID using the <in-grid> = int(<out-grid>) command.
Once both coverages are in vector polygon format, they can be merged using the UNION command. This creates a single coverage where each polygon contains an attribute from both parent coverages. Slivers less than the MMU are created in the merge, and can either be deleted in ARC/EDIT, or RESELECTed out when computing statistics.
Next, a cross-walk between the two vegetation classifications is created. The cross-walk attempts to provide a reasonable basis for comparison between to separate classification systems. The cross-walk used for this comparison is as follows:
Camp Williams Gap Analysis Cross-walk --------------------------------------------------------------------------------------- Oakbrush Oak Oakbrush Oakbrush, Sagebrush, Grass Mountain shrub --------------------------------------------------------------------------------------- Juniper Juniper Juniper --------------------------------------------------------------------------------------- Sagebrush Sagebrush Sagebrush Grass, Sparse shrub Sagebrush, Perennial grass --------------------------------------------------------------------------------------- Vegetated Agriculture Agriculture Agriculture Barren Agriculture --------------------------------------------------------------------------------------- Riparian Lowland riparian Riparian Highland riparian --------------------------------------------------------------------------------------- Barren or Annual weeds n/a Barren or Annual Weeds --------------------------------------------------------------------------------------- Urban Urban Urban --------------------------------------------------------------------------------------- n/a Salt Desert Scrub Salt Desert Scrub --------------------------------------------------------------------------------------- n/a Pinon-Juniper Pinon-Juniper ---------------------------------------------------------------------------------------
Coverage Comparison:
Now the coverages are ready to be compared. Two methods are incorporated
for the comparison: a qualitative method, and a quantitative method. The
qualitative method involves displaying each class on a map, and the
quantitative methods involves computing statistics.
For the qualitative comparison, nine maps are created - one for each cross-walk class. For each class, "xwalk1" is RESELECTed and drawn in white. The union coverage is then ASELECTed, and the identical class in "xwalk2" is RESELECTed and drawn in green. Next, the same class in "xwalk1" is again RESELECTed and drawn in red. Thus for each class, parent one only is drawn in white, parent two only is drawn in green, and areas in agreement are drawn in red. An example is shown here for the Oakbrush class. Williams oakbrush is shown in white, Gap Analysis oakbrush is shown in green, and areas of agreement are shown in red.
For the quantitative comparison, statistics are calculated in a similar way. For each RESELECTion, the STATISTICS command is run with the SUM AREA option, and the AREA is appended to a file. The file is then imported into a spreadsheet and three percentages are calculated for each class: "area both" divided by "area of parent 1", "area both" divided by "area of parent 2", and "area both" divided by "total area". The first two percentages show a percent agreement for each parent and the last percentage shows a percent agreement for both based on the total area classified by both. The following statistics were calculated for the comparison:
Class GAP Area (m2) Both/GAP WIL Area (m2) Both/WIL BOTH Area (m2) Both/Total 1 23594436.81093 0.828036087 32058602.69055 0.609416615 19537045.12651 0.540952713 2 9827524.503621 0.22897877 4469108.777856 0.503521974 2250294.474337 0.186803186 3 2486394.670363 0.625633514 3153334.058249 0.493310194 1555571.834741 0.380879549 5 52716703.80796 0.41290365 29130536.22549 0.747220004 21766919.40199 0.362296991 8 1886368.131069 0 95557.20440375 0 0 0 9 0 0 2792867.541231 0 0 0 12 958227.1636927 0.540317142 733641.875 0.705721116 517746.5625 0.440964697 32 1495989.31243 0 0 0 0 0 33 1439773.646896 0 0 0 0 0
Several generalization commands exist within Arc/Info (MERGE, GENERALIZE), however none of them incorporate ecological similarity in cover type. The routines simply merge a polygon with the neighbor that shares the longest border. The neighbor may or may not be a reasonable match. For example, a .5 hectare sagebrush polygon may share a border with a cloud polygon and a grass, sparse shrub polygon. If the cloud polygon shares a longer border, the sagebrush will be subsumed into the cloud polygon even though the cloud is an ecologically separate entity. By using the Smart-Raster-Eliminate program (Bassett, 1995) it is possible to weight class types so they are subsumed into ecologically related neighbors. The weighting matrix utilized was carefully evaluated by Tom Van Neil who performed the Camp Williams vegetation classification.
Cross-walking: Camp Williams and Gap Analysis utilized two separate classification conventions. The Camp Williams classification was primarily driven by the specific vegetation associations at the installation, whereas the Gap Analysis classification was labeled using the UNESCO system as required for National Gap Analysis standards. Because of the two systems, a direct cross-walk could not be established. Thus it was necessary to choose a grouping that made the most ecological sense. Grouping the classes becomes difficult when one class in the first system closely corresponds to two or more classes in the other system. In this case two possible solutions arise. The first is to cross-walk the one class to both of the other classes, and the second is to cross-walk the two other classes to the one class. As an example, Williams classifies agriculture into "Vegetated" and "Non-vegetated", whereas GAP only specifies "Agriculture". The first solution is to classify GAP "Agriculture" as both "Vegetated" and "Non-vegetated", whereas the second solution is to group the Williams "Vegetated" and "Non-vegetated" into one class - "Agriculture". The first solution is beneficial since it maintains the original number of input classes, but the result is a duplication in aerial extent. One class is used in two separate comparisons. The latter solution was chosen to eliminate this problem, but the tradeoff is a reduction in the number of classes since classes from both sides must be subsumed to establish a reasonable comparison. The cross-walk began with 10 Williams classes and 11 GAP classes, and ended with 9 cross-walk classes.
Comparisons: Two comparisons are chosen for optimum comparison of the datasets. Due to the spatial quality of the data, a purely statistical representation misses a great deal of spatial information, while a purely visual comparison lacks numerical precision. For the visual comparisons, a Venn diagram is the most intuitive. A Venn diagram comparison of two entities can be shown by utilizing only four colors. Shown below is a Venn diagram representation of the visual comparison. The two ovals represent each classification. White represents Camp Williams, green Gap Analysis, red both, and black neither.
The statistical comparison is also chosen for its simplicity. By necessity, it is assumed that the area identified by both classifications is correct. This assumption is realistic as two independent classifications are in agreement. However, it does not take into account temporal changes in land cover such that each classification may have correctly classified other areas but in the time between the classifications, the land cover may have changed. Thus the assumption is conservative in that it underrepresents the true accuracies. Three percentages are computed based on the areas of agreement: Both/GAP, Both/Williams, and Both/Total. All three percentages are measures of relative agreement. The first two measure how much GAP and Williams overclassified individually, and the third measures how much GAP and Williams overclassified together. It is important to remember that these percentages are relative to the area of agreement, and not necessarily relative to the actual spatial extent of the cover type.
Following is a brief display of the results of both comparisons for each of the nine cover types:
OakbrushPrinciple Components Classification
Map
Both/GAP .828 Both/Williams .609 Both/Total .541 Oakbrush shows the highest agreement. The Williams percentage is low due to a fire occurrence between the classification dates. Juniper
Map
Both/GAP .229 Both/Williams .504 Both/Total .187 Slightly better agreement could be obtained by including the GAP "Pinon-Juniper" class, but pinon-juniper does not exist within the Camp Williams boundary. Agriculture
Map
Both/GAP .626 Both/Williams .493 Both/Total .381 Sagebrush
Map
Both/GAP .413 Both/Williams .747 Both/Total .362 Sagebrush also shows a high agreement, and the GAP percentage is low due to the fire occurrence. (Fire burned sagebrush which was replaced with oakbrush.) Riparian
Map
Both/GAP .000 Both/Williams .000 Both/Total .000 Riparian areas were screen digitized for Gap Analysis separate from the unsupervised classification method. Riparian areas are highly dependent upon the date of the TM scene. 1988 (GAP) was a dry year, and 1993 (Williams) was a wet year. Bare ground, Annual Weeds
Map
Both/GAP .000 Both/Williams .000 Both/Total .000 This is a Williams class, and no corresponding GAP class that falls within the Camp Williams boundary applies. Urban
Map
Both/GAP .540 Both/Williams .706 Both/Total .441 Salt Desert Scrub
Map
Both/GAP .000 Both/Williams .000 Both/Total .000 Salt Desert Scrub classified by GAP does not exist within Camp Williams. Pinon-Juniper
Map
Both/GAP .000 Both/Williams .000 Both/Total .000 Pinon-juniper classified by GAP does not exist within Camp Williams.
The classes that were completely misclassified were generally of small spatial extent, however the riparian class should be viewed as highly suspect. Riparian areas are highly sensitive to damage, and the fact that no overlap occurred in this class is cause for concern. The GAP classification was performed during a dry year and showed more riparian areas than the Williams classification that was performed during a wet year. This is definately an area that requires further research.
The overlay of the principle components classification with the other datasets was very intriguing. The overlay added to the strength of the MMU misclassification hypothesis, but more surprisingly it showed the strong impact that classification date has upon the dataset. The area of largest descrepency in the comparison occurrs where a fire burned in between the two classifications.
Futher research is suggested to more closely define the effect of minimum mapping unit impact, and the accuracy of the riparian areas needs to be identified.
Van Neil, Tom. 1995. Vegetative Change Detection of Camp W.G. Williams, Utah Using Soil Adjusted Vegetation Indeces Derived From Landsat Satellite Imagery. Master's Thesis. Utah State University.