The chief objective of this study was to evaluate the airborne AISA hyperspectral scanner for identifying and mapping tidal wetland vegetation. The high spatial resolution (1 to 2 m) of which the instrument is capable is clearly useful in the wetland environment. The two meter pixel size employed here is well matched to the scale of such wetland features as the drainage system and the vegetation communities.
Two problems with this AISA digital data set must be noted. The sensitivity of the instrument for wavelengths shorter than about 500 nm is low. Compensation in this spectral region by increasing gain results in a noticeable increase in signal noise. This most strongly affects the blue bands (1-5) of this data set. Increased sensitivity in the blue portion of the spectrum might improve the ability of the reflectance data to discriminate between vegetation types. Secondly, an asymmetrical "rolloff" (or decrease) of reflectance in spectral bands 1 through 10 over about one quarter of the scan introduces a small variable bias in the image data which may increase difficulties in classification. This problem, which appears to be of instrumental origin, should be corrected in the instrument or adjusted for prior to delivery of the data for image analysis.
This flight took place later in the season than optimum for obtaining spectral signatures from growing marsh vegetation. The two month period July and August, perhaps including early September, is the time of peak green biomass for tide marsh vegetation In late September the live biomass of the salt marsh vegetation was undergoing senescence as the plants prepared for winter dormancy. At this time different vegetation stands would show more diversity in canopy vigor than during the peak growing season resulting in greater diversity in spectral reflectance than during the peak season. This increases the difficulty in classifying the vegetation cover using its spectral reflectance alone.
It was found necessary for classification to segment the image into wetland and upland areas for separate classification. The wetland area had to be separated from upland forest manually because of spectral confusion between forest shadow and marsh vegetation.
The length of time required to complete the overflight of the Milford Neck study area was four hours. During this time the solar illumination and shadow patterns change sufficiently between the portion flown before local noon and that flown after local noon as to affect the spectral identification of similar vegetation stands in the two parts of the image. They may not be recognized as similar. This might require breaking the image into appropriate regions for classification.
All of the above factors may contribute to explaining why the unsupervised approach to classifying wetland tide marsh vegetation in this image was not successful.
We had more success using a supervised approach to classification in which a number of training areas for each vegetation type were located in different parts of the image. The training sites showed good spectral separation between vegetation types. However, in the supervised classification itself there was considerable confusion among vegetation classes, especially between Phragmites and higher density canopies of S. alterniflora. In preparing the final classification it was necessary to manually reclassify a number of cases where these two classes were confused.
An evaluation of the ability of spectral criteria in this supervised classification to identify pixel size units of the wetland land cover classes showed an overall users accuracy of 76%. Users accuracy for the combined S. alterniflora classes, the combined Phragmites classes, and salt hay class also were all 76%. Users accuracy for high tide bush was 95%. Users accuracy for the combined shallow water/tidal flat category was 74%. These levels of accuracy may be graded as fair if an 80 or 85% success rate would be considered good. However, given the difficulty of this spectral separation problem, these result are quite acceptable.
We could not verify through these ground observations a reliable ability to spectrally distinguish shallow water/tidal flat from open water although the spectral signatures of these two are quite distinct. The reason is probably due to differences in the tide level and algal growth between time of image and ground observation. Significant differences in vegetation extent between the time of image and ground visitation in the poorly drained Phragmites management area east of the canal near Big Stone Beach probably contribute to this issue.
A low level of classification accuracy was found for individual stands of living Phragmites and areas of Phragmites canes. The two tended to be confused partly because of problem of vegetation regrowth between time of image and time of ground visitation. There is also difficulty identifying in the image low density Phragmites mixed with S. alterniflora. A tendency to confuse Phragmites and high canopy density S. alterniflora is also noted. Combining the two Phragmites classes significantly increased their combined classification accuracy to 76%.
The combined classes of high and low canopy density S. alterniflora showed a users accuracy of 76% while the accuracies of the individual cases were quite low largely because of confusion with one another. A rigorous field definition of these two cases would have to be developed to allow the field verification of these two spectral classes.
Salt hay was well recognized in the classification when it occupies many contiguous pixels and is lodged. However, when growing in a bunched, hummocky, habit it is not lodged. In the classification is then easily confused with S. alterniflora. This result is reflected in the lower producers accuracy of 63% for this class although the users accuracy for salt hay is 76%
High tide bush appears to be well identified. This class showed a producers accuracy of 72% and a users accuracy of 95%.
In general, use of the hyperspectral AISA imager did not allow for as accurate a spectral separation of dense canopies of S. alterniflora and Phragmites as we expected to achieve. Factors contributing to this are believed to be 1. the lateness in the season (progress toward senescence), and the dryness of the prior growing season leading to different levels of stress in canopies from place to place; 2. Differences in ground illumination across the scene due to the long time required to image the whole study area; 3. Certain problems of uniformity and limitations of sensitivity of the AISA digital data toward the shorter wavelengths, especially shorter than 500 nm. An instrument with greater sensitivity in the blue spectral regions and which extends into the mid infrared may improve the identification of this wetland vegetation. These results, especially the high users classification accuracy obtained for high tide bush, suggest that higher levels of accuracy for classification of tidal wetland vegetation here may be attained with further improvements in airborne spectral scanning instrumentation.