Classification Results

The output error rates computed on the test data sets for each Level-3 pixel-based classifier and "spatial" classifier are listed in Tables 3 and 4 below. The apparent ranking of the pixel-based classifiers (based solely on the classifier error rates computed from the test data sets) from best to worst are NN-MNF13, MLC-CA, MLC-DBFE, NN-5, and MLC-5. However, based on visual inspection of the entire scene, the classification of the entire data set results in a ranking from best to worst is NN-MNF13, MLC-5, MLC-CA, MLC-DBFE and NN-5. Performance of the contextual and region based classifiers ranks them as MRF-MNF13 and MRF-PCA8. The results of each classifier are illustrated in the sections below.

Table 3. Pixel-based Classifier Accuracy
Vegetation Type NN-5 MLC-5 NN-MNF MLC-DBFE MLC-LDFE
Scrub 95.3 % 84.8 % 97.2 % 81.3 % 80.2 %
Willow Swamp 93.8 % 90.5 % 96.3 % 87.2 % 88.5 %
Cabbage Palm Hammock 87.1 % 87.9 % 88.2 % 82.8 % 93.8 %
CP/Oak Hammock 63.1 % 63.1 % 83.0 % 53.4 % 94.4 %
Slash Pine 62.3 % 67.1 % 83.0 % 53.4 % 94.4 %
Oak/Broadleaf Hammock 46.0 % 47.2 % 90.8 % 98.3 % 83.8 %
Hardwood Swamp 88.6 % 94.3 % 100.0 % 69.5 % 89.5 %
Overall Uplands 76.6 % 76.4 % 92.5 % 78.0 % 86.8 %
Graminoid Marsh 74.8 % 74.2 % 98.6 % 80.3 % 79.1 %
Spartina Marsh 87.3 % 89.6 % 90.8 % 77.9 % 83.5 %
Cattail Marsh 83.6 % 90.8 % 94.8 % 75.7 % 82.2 %
Salt Marsh 97.1 % 96.9 % 99.3 % 88.1 % 87.4 %
Mud Flats 92.8 % 73.6 % 83.8 % 79.9 % 82.7 %
Overall Wetlands 87.1 % 85.0 % 93.5 % 80.4 % 83.0 %


Table 4. Contextual-based Classifier Accuracy
Vegetation Type MRF-PC8 MRF-MNF13
Scrub 93.3 % 93.4 %
Willow Swamp 90.1 % 93.0 %
Cabbage Palm Hammock 89.5 % 93.4 %
CP/Oak Hammock 74.6 % 83.7 %
Slash Pine 94.4 % 78.3 %
Oak/Broadleaf Hammock 96.9 % 92.6 %
Hardwood Swamp 96.2 % 98.1 %
Overall Uplands 96.2 % 98.1 %
Graminoid Marsh 81.9 % 82.6 %
Spartina Marsh 91.5 % 87.3 %
Cattail Marsh 95.0 % 95.0 %
Salt Marsh 98.3 % 95.2 %
Mud Flats 79.5 % 79.7 %
Overall Wetlands 89.2 % 87.9 %


Output Images from tested classifiers

Pixel-based Classifiers
Neural Net - 5 band selection
Maximum Likelihood - 5 band selection
Neural Net - 13 MNF Bands
Maximum Likelihood - Decision Boundary Feature Extraction
Maximum Likelihood - Linear Discriminant Feature Extraction

Contextual Region-based Classifiers
Markov Random Field - 13 MNF Bands
Markov Random Field - 8 PC Bands


Discussion of Results

Overall, the contextual-based classifiers seemed to do a better job of classification than the pixel-based classifiers. The Markov Random Field classifier using the MNF input data most accurately classified the overall image and had an upland accuracy of 90.3% and a wetland accuracy of 87.9%. The MRF-MNF13 result seemed to accurately map the distribution of scrub as well as correctly identify many of the upland groupings. The wetland vegetation (both within and outside the impoundments) was also mapped well with the only exception being an over-classification of salt marsh at the top of the image.

The MRF-PCA8 also performed well in classifying the entire image, but the amount of scrub lying to the north of the road was underestimated. The MRF-PCA8 also misclassified willow swamp and cabbage palm hammock in impoundment T24-D as slash pine. However, this input combination did better than the MRF-MNF13 for mapping salt marsh. Both the MRF-MNF13 and the MNF-PCA8 provided good overall classifications, but the MRF-MNF13 was better overall and would require less post-processing to import the result into a GIS system.

Of the pixel-based classifiers, the NN using the MNF data performed the well with an upland accuracy of 92.5% and a wetland accuracy of 93.5% based on training data. Although results of the classifier applied to the entire image were quite good, some of the wetlands were classified to have too much salt marsh. Since both the contextual and the pixel-based classifiers performed well using the MNF data as inputs, it is assumed that this transformation can be used successfully for the classification of hyperspectral imagery.

The Maximum Likelihood classifier using the linear discriminant analysis (Canonical Analysis) actually yielded good results for the entire data set, although the error accuracy for the uplands is 86.3% and a wetlands is 83.0%. LDFE was able to separate the upland classes well, with the only notable exception being an over-classification of oak hammock. However, this method was able to accurately identify several small regions of slash pine which other classifiers incorrectly labeled. Unfortunately, however, this classifier misclassified Spartina marsh as salt marsh in many of the impoundments (T10-F through T10-I).

The ML classifier had an upland accuracy of 76.4% and a wetland accuracy of 85.0%. However, visual evaluation of the classification results for the entire data set looked much better than the error rates computed on the test data sets. While the ML classifier performed well on the full data set, it had difficulties in identifying slash pine communities and misclassified cabbage palm hammock in isolated regions of the impoundments.

Using the decision boundary feature extraction method, the ML classifier overall had accuracy of 78.0% for uplands and 80.5% for wetlands, and the overall results did not look as good as results from other classification methods. The MLC-DBFE overclassified salt marsh in both the T10-D through T10-I impoundments as well as in impoundment T24-D and it underclassified scrub.

The NN-5 classifier yielded an accuracy similar to those obtained by the ML classifier for the test data sets, but visual inspection of the full data set showed that this classification was not as good as the other classifier/inputs. Analysis of the full data set showed that the NN-5 classifier had difficulty in the identification of slash pine communities, misclassification of some portions of spartina marsh and misclassification of isolated cabbage palm hammocks in the impoundments.