IMAGE ANALYSIS AND MAPPING
(click on any thumbnail to show full-sized image in another window)

(C) Annotated, pseudo-color infrared image of the Elkhorn Slough study area created from AVIRIS bands 41, 28, 18 (RGB=721nm, 656nm, 557nm). The waters of Monterey Bay, Elkhorn Slough and Moro Cojo Slough show up as dark areas because near infrared wavelengths (~754nm) are absorbed in the very upper surface of water. The red areas indicate vegetation, which reflects more in the near infrared than in green visible wavelengths. The next steps in computer processing of the data included corrections for atmospheric conditions using ATREM, an algorithm developed by the University of Colorado - Boulder. All subsequent computer analysis was done using ENVI software, including the Minimum Noise Fraction (similar to a Principal Component Analysis) to segregate noise and determine the inherent dimensionality of the data.

(D) Greyscale image of band 68 (1014 nm). Because one dataset did not cover the entire area of interest, two AVIRIS scenes from different flight lines on the same date were mosaicked together, and then a spatial subset was taken to outline the study area and reduce the dataset. The images were then converted from radiance to apparent reflectance with ATREM (ATmospheric REMoval program, University of Colorado - Boulder). However, cirrus clouds visible in the image still present a problem, and along with the low SNR of 1992 AVIRIS data, have complicated subsequent computer analysis of the imagery.

A principle components transform, known as the Minimum Noise Fraction (MNF) algorithm, was run to reduce spectral noise and dimensionally subset the data. Due to the low SNR of the data, only the first 30 of 224 possible bands were selected for further analysis. A Pixel Purity Index (PPI) was run on these spectrally coherent bands to determine the "purest pixels". The result of this may be seen in (E), where the "purest" pixels are shown in white.

After initial post-processing of the data, (F) shows an unsupervised classification of the "purest" pixels was successful in mapping ground cover types, including a vegetation class where Salicornia sp. may be found. Georectification of the image, and subsequent ground-truthing of these areas in the field, resulted in 90% accuracy.

(G) 5-D example of ENVI's n-Dimensional Vizualization of the "purest" pixels showing colored clusters of pixels which may be interpreted as spectral endmembers.

(H) shows the colored spectral endmember pixels plotted back into their correct geographic location, and overlayed on a greyscale image of MNF band 2. In addition to the grouped clusters shown in the above 5-D Visualization image, further analysis enabled the extraction of additional endmembers shown here.

Unfortunately, Salicornia sp., the initial focus of this study, was not able to be spectrally extracted using this method. This may be due to 1) the 20m spatial resolution of existing AVIRIS imagery is too large for the small-scale patchiness inherent to Elkhorn Slough wetlands; and 2) the above mentioned low SNR of the 1992 AVIRIS dataset compounded with clouds and moisture in the atmosphere.

(I) Mapped results from ENVI's supervised Spectral Angle Mapper (SAM) algorithm using only the "pure" pixels and 8 user-defined training classes chosen from a priori knowledge of the field area. Differences from the unsupervised classification include a small portion of the evaporite pond being classified as asphalt/dirt, and a large portion of the Eucalyptus grove being mapped as water. Although exact conditions of the field area in 1992 (when the imagery was acquired) are not known, these classifications are unlikely.

(J) Mapped results from the SAM algorithm, using all pixels in the image (not just the "pure" pixels). The algorithm was run with 1 user-defined training class: wetland vegetation (including Salicornia sp.).

The images shown above are examples of some of the maps created from this study. The final products are digital, georeferenced maps, available for printing as hard-copies or for importing into a Geographic Information System (GIS) software package for future use.

RATIO ANALYSIS
(click on any thumbnail to show full-sized image in another window)

K
(K) shows a series of signatures collected along a transect of Salicornia sp. in Elkhorn Slough on 14 Aug 98. The signatures range from healthy "green" vegetation to an unhealthy "grayish" dying plant. Note how the healthy signal shows a high green reflectance (~550nm), a relatively strong red absorption feature (~650nm), and very high reflectance in the near infrared (NIR) plateau (~750-925nm). As the plants die, they lose their green reflectance, red absorption and high NIR plateau, and the overall signature geometry "flattens out".

L M
(L) shows two standard vegetation indices (NDVI and SAVI), commonly used as indicators of health, several stress ratios, and the Red edge Vegetation Stress Index (RVSI) applied to the data collected along the Salicornia sp. transect shown in (K). The samples range from healthy (1) to poor (7). The NDVI and SAVI show similar trends along the transect, with the four healthiest plants having very similar values and then declining through the three "stressed" plants. Three of the stress ratios (695nm/760nm, 710nm/760nm, and 695nm/850nm) are consistent with the known trend of health along the transect. However, both the 695nm/420nm stress ratio and the RVSI have different patterns from these trends (M).

N
While the 695nm/760nm ratio shows little variation and no trend along the transect (N), there is a strong and very similar increase in the other two stress ratios. The 710nm/760nm ratio varies a factor of 2 (from approx 0.2 to 0.5), but the 695nm/850nm ratio varies a factor of 5 (from approx 0.05 to 0.25) and was therefore used for further analysis in this study.

(O) shows the 695nm/850nm stress ratio applied to the pixels identified as wetland by the analysis shown in (J). Pixels shown in red have higher ratio values and indicate areas of potential ecosystem stress at the time of the AVIRIS data collection. Although many of these areas may not have appeared visibly stressed to a casual observer, these examples illustrate the ability of hyperspectral imaging spectroscopy to detect physiological changes in ecosystems before it becomes apparent to the human eye, leading to the potential of arresting problems before they go unchecked. Using maps such as these, environmental managers can focus their investigations and programs on specific locations. When applied to multi-temporal data, seasonal patterns and early detection of potential problem areas may be tracked and monitored.

Conclusions

Good results were obtained from ENVI's unsupervised classification technique to create a basic ground cover map. Reasonable results were obtained using the Spectral Angle Mapper (SAM) supervised classification technique with one user-defined training class to map the extent of wetland vegetation. However, the SAM was not able to identify substrate types when compared to library spectra collected in the field.

Of the five stress ratios compared in this study, the 695nm/850nm stress ratio provided the most consistent results for ecosystem health analysis in the wetland environment of Elkhorn Slough, while neither the 695nm/420nm ratio nor the Red-edge Vegetation Stress Index (RVSI) detected gradients of physiological stress.

Sources of error included the 1) inability of ATREM to account for temporally and spatially varied water vapor concentrations; 2) ATREM's failure to correct for cloud scattering effects.

This investigation found the poor signal to noise ratio in the 1992 AVIRIS data limited spectral identification of endmembers, and the 20 m ground resolution pixel size minimized the ability to detect small-scale patchiness in this wetland ecosystem.

Hyperspectral techniques can provide rapid results, virtually in real-time, to evaluate research and monitoring efforts in coastal environments, to detect and evaluate environmental changes, and to detect, monitor and interpret biological responses to those changes. This technology will provide a powerful tool for assessment of current management programs in Elkhorn Slough and may be easily transposed to other coastal environments.

Future studies must give priority to ground cell resolution when planning the use of remote sensing techniques for mapping and monitoring.

Acknowledgements

This presentation, a culmination of studies from my M.S. thesis, Hyperspectral Imaging Techniques Applied To Ecosystem Health In Elkhorn Slough, CA (1999), at UC Santa Cruz, was conducted as part of a joint investigation with Lawrence Livermore National Laboratory (LLNL) and supplemented with funding from the Cooperative Institute for Coastal and Estuarine Environmental Technology (CICEET). I also wish to thank UCSC's ReEntry Services, the American Museum of Natural History, and the Houston Underwater Club/Seaspace for their additional financial support. Special thanks goes to Brigette Martini, Bill Pickles, Don Potts, and Daria Siciliano for help in the field.