Table 1 Comparison of Classification and Model Results for Two Study Sites
Accuracy Sm
Method Site Forested | Other | Forested | Other Overall Overall |
Wetland | Wetland Upland Upland Kappa
Acadia user 0.70 0.50 0.74 0.72 74% 0.46
Unsuper- (312) producer 0.12 0.06 0.96 0.85
vised Orono user 0.58 0.62 0.81 0.81 72% 0.60
(431) producer 0.71 0.42 0.73 0.82.
Acadia user 0.41 0.41 0.81 0.94 75% 0.53
Tasseled (309) producer 0.25 0.75 0.91 0.72
Cap Orono user 0.64 0.41 0.84 0.84 74% 063 |
(362) producer 0.65 0.47 0.73 0.97
Acadia user 0.80 0.41 0.84 0.91 78% 0.61
Hybrid (282) producer 0.35 0.91 0.95 0.56
Orono user 0.75 0.82 0.80 0.86 81% 972 73
(253) producer 0.57 0.64 0.87 0.98
Existing GIS Orono user 0.82 0.64 0.83 0.80 80% 0.71
Model (406) producer 0.66 0.71 0.87 0.92
Acadia user 0.62 0.90 0.90 0.81 82% 0.70
Integrated (309)” producer 0.77 0.45 0.86 0.89
(Revised) Orono user 0.78 0.67 0.86 0.80 81% 0.72
Model (402) producer 0.78 0.70 0.87 0.76
Notes: The classification results of Orono were derived from Ahl (1994) except the result of the integrated model.
"Sample size is in parenthesis.
4.3 Integrated Model
The integrated model combined one classified image (forest
recoded as 16 and non-forest recoded as 0) and four weighted
GIS layers in a series of 2%”, where k indicated the order of layer.
An analysis image with 32 combination levels were created by
map algebra using an additive approach. An analysis matrix was
generated (Table 2). The columns represent the reference data in
four classes which were separated by row into two primary
groups of forest and non-forest. The rows represented the 32
combination levels and indicated the distribution of reference
data within levels. The aggregation of 32 combination levels into
four classes was defined by the accumulated ratio and controlled
by the balanced ratio.
In the non-forest group, the threshold for other wetland and other
upland is obvious (Table 2), even the balance ratios, 0.48 and
0.58, are not comparable. It indicated that combination level 0
was a source of confusion and a poor outcome for the other
wetland class as expected. On the other hand, the threshold for
forested wetland and forested upland was set between
combination level 20 and 21. It was observed that increasing the
ratio of forested wetland will dramatically drop the ratio of
forested upland. Besides, the final ratios would reflect in the
producer agreement of the confusion matrix except in the other
upland category ( Table 1 ). The producer and user accuracy for
forested wetland were 0.77 and 0.62 respectively. The overall
agreement was 82% and the Kappa was 0.70. These results are
the best among the two and four classes approach conducted in
the Acadia study site.
International Archives of Photogrammetry and
4.4 Integrated Model Applied in the Orono Study Site
The same algorithm was applied using the Orono data set (Al,
1994). First, the slope percent layer was redefined by examining
an analysis matrix. Second, individual GIS layers were evaluated
by cross tabulation. The contribution of four GIS layers was the
same as defined in the existing model (Ahl 1994). They were
NWI, hydric soil, slope, and stream buffer layer. Similarly to
Acadia, an analysis matrix can be generated for the four groups
aggregation analysis. The final ratios reveal a homogeneous
confusion matrix. Forested wetland producer and user accuracy
were both at 0.78, and the overall agreement and the Kappa were
81% and 0.72 respectively. These results are the best among al
the approaches that had been conducted in Orono study sit
(Table 1).
The integrated model improved the results at Orono by
equalizing the accuracy between user and producer agreement.
This can be observed by comparing the Orono integrated mod!
results with other approaches (Table 1). This emphasizes the
utility of the analysis matrix and balance ratio in the aggregation
process. Beyond the majority editing rule, an expected result
might be revealed through examining the final balance ratio. The
integrated model makes both overall accuracy and Kappa
correction more comparable at two study sites (Table 1). This
result suggests that the integrated model was less influenced by
the size of study area. Moreover, the significance test of Kap!
correction performed consistently in both study sites when the
integrated model was involved (Table 1). Although the integrated
model provided the best result among all experiments on both
study sites, it was not determined to be a significan
improvement over the conventional hybrid approach.
422
Remote Sensing. Vol. XXXI, Part B7. Vienna 1996