Mapping of Coastal Wetlands via Hyperspectral AVIRIS Data

Amy L. Neuenschwander ⁽¹⁾, Melba M. Crawford ⁽¹⁾, and Mark J. Provancha ⁽²⁾
(1) Center for Space Research, University of Texas, 3925 W. Braker Lane, Suite 200, Austin, TX 78759
(512) 471-5573, amy@csr.utexas.edu
(2) Dynamac Corporation, Mail Code: Dyn-2, Kennedy Space Center, FL 32899
(407) 853-3281

Abstract - The wetlands located on the west shore of the Kennedy Space Center (KSC) and the Indian River are critical habitat for several species of water fowl and aquatic life. Mapping of the land cover and its response to wetland management practices using remotely sensed data from a variety of sensors is the focus of a multi-year project. In 1996, hyperspectral data were acquired using the Airborne Visible Infrared Imaging Spectrometer (AVIRIS) over the KSC complex. The test site for this study consists of a series of impounded marshes with vegetation communities ranging from low, halophyte marshes to high, graminoid savannas to forested wetlands. The capability of the AVIRIS data for discriminating the various wetland communities is investigated herein. Preliminary results indicate that the improved spectral resolution of the AVIRIS sensor is advantageous in areas such as wetlands where much of the vegetation is spectrally similar.

INTRODUCTION
The use of impoundments and other structural marsh management methods within wetlands for mosquito control and other wetland uses has led to considerable discussion regarding the potential effects on the natural marsh sustaining processes and the emergent marsh vegetation [1]. As a result, many impounded marshes are currently being reconnected with the adjacent estuaries so that they may return to a more natural hydrology. Thus, there is a need to develop a protocol which can quickly and accurately classify and map the distribution of vegetation types within the impoundments using remotely sensed data acquired over a multi-year time horizon.

The test site for this research consists of a series of impounded estuarine wetlands of the northern Indian River Lagoon (IRL) which reside on the western shore of the Kennedy Space Center. The impoundments were created during the 1950's and 1960's for the purpose of mosquito control. The marshes along the IRL contain both high and low marsh communities. The three dominant marsh groups which comprise the high marsh communities are cabbage palm savanna, sand cordgrass, and black rush. The cabbage palm savanna consists of isolated canopies of Cabbage Palm (Sabal palmetto) and a graminoid layer of sand cordgrass (Spartina bakerii ), whereas sand cordgrass marsh is dominated by Spartina bakerii and black rush marsh (Juncus roemerianus) is dominated by Juncus. The low marsh communities are dominated by salt tolerant grasses and halophytes. The primary salt tolerant grass is Distichilis spicata. Halophytes typically include Batis maritima and Salicornia virginica [2].

This study also includes investigation of upland vegetation as it is adjacent to the impounded wetlands, so the two units are related hydrologically. In addition, accurate classification and mapping of upland vegetation is important for monitoring habitat of endangered birds. The majority of the upland vegetation at KSC is oak scrub and saw palmetto scrub. Other upland communities include slash pine (Pinus elliottii) and hardwood swamps which are dominated by deciduous trees such as Red maple (Acer rubrum). Dense hammocks of Cabbage Palm (Sabal palmetto) and Live Oaks (Quercus virginiana) are also common.

REMOTELY SENSED DATA

The NASA AVIRIS spectrometer acquired data over the Kennedy Space Center, Florida on March 23, 1996. The AVIRIS sensor collects 224 bands of 10 nm width in the reflected visible and near infrared spectrum (400 - 2500 nm) with a spatial resolution of 18 m. The AVIRIS data were converted to apparent surface reflectance using the ATREM [3] (Atmosphere Removal Program) developed by the Center for the Study of Earth from Space at the University of Colorado, Boulder, Co.

A Maximum Noise Fraction Transformation (MNF) [4] was performed on the ATREM corrected data to reduce the dimensionality of the data set. The MNF transform selects the ith linear combination which maximizes the noise fraction of all linear transformations orthogonal to the previously computed j=1,..i-1 transformed bands, thereby isolating noise in the residual band (which can be filtered). Application of the procedure maximizes the signal to noise ratio of successive combinations and eliminates redundant information between highly correlated bands, thereby providing a subset selection procedure for hyperspectral data.

CLASSIFICATION METHODOLOGY
The proposed hierarchical architecture for classification of a coastal environment entailed dividing the classification process into simpler tasks based upon class separability. The three level hierarchical classifier shown in Figure 1 was employed for the KSC test site. The Level-1 classifier separated the image into water and land, and the Level-2 classifier discriminated between the land classes consisting of wetlands and uplands. At Level-3, various algorithms were investigated to separate the land cover into the "final" classes. Classification accuracy was improved by using the hierarchical architecture relative to a single stage classification approach.

Figure 1. Schematic Diagram of Hierarchical Classification of AVIRIS data.

The Level-1 classifier used a simple isoclustering method. A biomass index similar to a NDVI was employed at Level 2 to separate the uplands from the wetland vegetation. Because they have much more biomass than marshes, the willow swamps and hardwood swamps are included in the "uplands" category, even though they are actually wetland communities. Both a Gaussian Maximum Likelihood (ML) classifier and a neural network (NN) with multiple combinations of inputs were investigated at Level 3. A multi-layer perceptron neural network with one hidden layer and a scaled conjugate gradient training algorithm was applied to various combinations of input data. The Maximum Likelihood (ML) classifier was used primarily as a comparative approach.

Description of Inputs
Selected bands of the original ATREM corrected data and MNF transformed bands computed from the complete data set were analyzed. The ML and NN classifiers were first applied to a data set containing five bands which corresponded to significant portions of the spectrum for vegetative response: Band 34 (672 nm center), Band 42 (739 nm center), Band 64 (960 nm center), Band 118 (1462 nm center), and Band 138 (1661 nm center). The NN classifier was also run on a data set computed using the last 13 eigenvectors of the MNF transformation.

Description of Outputs
Training sites for each class were determined using previous classification maps, aerial photography, and extensive knowledge of the area. The initial "training" data were randomly divided into training and test data sets with equal numbers of observations.

The output classes for the uplands were scrub, willow swamp, cabbage palm hammock, cabbage palm/oak hammock, slash pine, oak/broadleaf hammock, and hardwood swamp. Wetland classes were selected to be graminoid marsh, Spartina bakerii marsh, cattail marsh, salt marsh, and mud flats. The graminoid marsh class is actually a mixture of marsh grasses which do not appear in large homogeneous, spatially contiguous groups that can be readily identified in the imagery.

CLASSIFICATION RESULTS
The output error rates computed on the test data sets for each Level-3 classifier are listed in Table 1. The apparent ranking of the classifiers (based solely on the classifier error rates computed from the test data sets) from best to worst are NN-MNF13, NN-5, and MLC-5. However, based on visual analysis, the classification of the entire data set results in a ranking from best to worst of NN-MNF13, MLC-5, and NN-5.

Overall, the NN using the MNF data performed the best with an upland accuracy of 92.5% and a wetlandaccuracy of 93.5% based on training data. Although results of the classifier applied to the entire image were quite good, some of the impoundments were classified to have too much salt marsh.

Table 1: Classifier Error Rates

	Classifier Type
Vegetation Type	NN 5	MLC 5	NN MNF 13
Scrub	95.3	84.8	97.2
Willow Swamp	93.8	90.5	96.3
Cabbage Palm Hammock	87.1	87.9	88.2
CP/Oak Hammock	63.1	63.1	91.7
Slash Pine	62.3	67.1	83.0
Oak/Broadleaf Hammock	46.0	47.2	90.8
Hardwood Swamp	88.6	94.3	100.0
Overall Uplands	76.6	76.4	92.5

Graminoid Marsh	74.8	74.2	98.6
Spartina Marsh	87.3	89.6	90.8
Cattail Marsh	83.6	90.8	94.8
Salt Marsh	97.1	96.9	99.3
Mud Flats	92.8	73.6	83.8
Overall Wetlands	87.1	85.0	93.5

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.

Overall, the NN-5 classifier yielded similar accuracies to those computed for 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.

CONCLUSIONS
Mapping of the land cover and its response to wetland management practices using remotely sensed data from the AVIRIS sensor is being investigated at the KSC complex. Relative to previous studies of the area based on Landsat TM, the increased spectral resolution of AVIRIS data is useful for the identification of wetland vegetation (all in the 90% range) as well as for discrimination between various upland vegetation types.

Figure 2. Classified image of western shore of KSC using NN-MNF13 classifier.

REFERENCES

EPA, "Ecological impacts and evaluation criteria for the use of structures in marsh management", EPA-SAB-EPEC-98-003, January, 1998.
Schmalzer, P.A., "Biodiversity of saline and brackish marshes of the Indian River Lagoon: Historic and current patterns", Bulletin of Marine Science, 57(1):37-48, 1995.
CSES (Center for the Study of Earth from Space), ATREM (Atmosphere Removal Program) Users Guide, University of Colorado, 1992.
Green, A.A., M. Berman, P. Switzer, and M.D. Craig, "A Transformation for Ordering Multispectral Data in Terms of Image Quality with Implications for Noise Removal," IEEE Trans. on Geoscience and Remote Sensing, 26(1): 65-74, 1988.