Archived: July 29, 2004



Classification Methods
Our initial classification will be focused primarily upon the vegetation (i.e. no industrial) of the Rincon Bayou. Due to the cloud cover, the prelimanary investigation of CAMS and AIRSAR covers the western portion of the bayou and does not include salt marshes, therefore they will not be included into the initial classification results. Visual interpretation of the CAMS and AIRSAR imagery yields the following vegetative classes that should be readily identifiable for the initial study:

For more information about the wetland and marsh definitions, see wetland.

For photos of the vegetation in the Rincon Bayou, see photos.


The CAMS imagery that was acquired over the Rincon Bayou required two flight paths for total coverage. Therefore, the two adjacent strips were mosaicked for image analysis. The difference in the radiometric properties of each CAMS strip is apparent. To reduce this effect, a histogram matching algorithm was applied to the strips prior to mosaicking. The histogram matching, however, did not totally remove the radiometric differences.

A graphical mask was created for classification of only the Rincon Bayou. Due to the large amounts of clouds present in the CAMS imagery, an additional graphic mask representing the clouds and cloud shadows were also created as to prevent errors in classification.

Vegetation mapping of the Rincon Bayou was performed using two classification algorithms, maximum likelihood and neural nets. Both algorithms are supervised methods and require training for proper classification. The training data was selected by choosing regions in the imagery which corresponded to a 1992 wetlands map created by the U.S. Bureau of Reclaimation.

Classification Results
The best results for vegetation mapping required using both the CAMS and the AIRSAR imagery. Classification signatures were created for CAMS channels 1,4 thru 9 and all of the AIRSAR channels. The classification using these channels resulted in an excellent preliminary result of the wetland vegetation.

Maximum Likelihood Classifier
Maximum Likelihood classification results of the Rincon Bayou proved to be quite comparable with the wetlands maps generated in 1992. The classifier had some difficulties in separating fresh and brackish marsh. This is due to the common vegetation found in varying salinity water. Further refinement of the classifier should yield excellent results.


Neural Net Classifier
Neural net classification results based upon the classifier from PCI did not work quite as well as the maximum likelihood. There was a significant overcompensation of low brackish marsh in areas that are known transitional and upland areas. In addition, some of the uplands were misclassified as high freshwater marsh. However, other neural net algorithms may prove successful in classification and will be investigated for this application.




Error Analysis
Error Matrices were generated for both classification algorithms as an indicator of successful classification. The error statistics were computed on the training data. As Table 1 below indicates, the maximum likelihood classifier did well with the water, mud flats, and the woodland areas. The largest discrepancies occur between the fresh water marsh and the transitional areas. This is partly due to the fact that many common wetland species can occur in both fresh and brackish conditions.


Table 1: Maximum Likelihood Classification Error Matrix



Buttons

Sunday, 01-Aug-2004 00:24:38 CDT
CSR/TSGC Team Web