Introduction

Remote sensing offers the potential to observe the responses of coral reef ecosystems to ecological and environmental perturbations on a geographical scale not previously accessible. However, coral reef environments are optically, spatially, and temporally complex environments which present complex challenges for extracting meaningful information concerning the ecology and vitality of coral reef communities. One of the keys to this remote sensing problem of coral reef "health" is the development of a methodology that relates the remotely sensed satellite signal to the optical signal generated by coral reef environments or reef bio-optics. Using a bio-optical approach, it may be feasible to remotely map benthic habitats. Traditional marine science has always utilized in-situ monitoring to detect community change, but the archival record of satellite-based information contains synoptic mesoscale information that could also be tapped to examine change in the coral reef ecosystems. This approach becomes more attractive when one considers the remoteness of most reefs and the expense of expeditionary travel. Space-based observations are probably the most cost-effective way to observe temporal changes in coral communities. More simply stated, the question becomes, "Can remote sensing be used to follow changes in coral reef biodiversity?"

Scientists do not yet know definitively what is causing the decline of some of the reefs in Caribbean, it is suspected by many researchers that multiple stressors including direct and indirect anthropogenic impacts are at the root of the problem. Furthermore, in South Florida, the situation is unique, in that we seem to be observing the effects of environmental impacts cascading through multiple ecosystems. The Florida Reef Tract is at the downstream end in the hydrology of South Florida, a system that has experienced explosive human population and agricultural growth in the last 50 years. The water of the Gulf Stream at its seaward margin also incorporates materials from elsewhere in the Caribbean and Gulf of Mexico, which may contribute additional stress to the system.

We focused on Landsat TM imagery to investigate reef change. This instrument has spatial resolution of 30 meters and three visible broad-spectrum bands (red, green, and blue). This provides one of the only available remote sensing platforms for locating reefs, mapping their general zonational patterns and dominant benthic communities (coral, seagrasses, sand, etc.) (Biña, et. al., 1978; Jupp, 1986; Manier and Jaubert, 1985; Kuchler, et. al., 1988; Boor and Pichon, 1997; Mumby et. al., 1998; Dustan et. al., 1998 in press).

While there are better sensors in development and slated for launch in the coming years, Landsat TM offers the potential for time series observations. The first Thematic Mapper instrument was sent into orbit aboard Landsat 3 in July of 1982 and began to capture imagery shortly after launch. It was operated by NASA’s Goddard Space Flight Center (GSFC) for two years and, in January 1983, operations of the Landsat system were transferred to the National Oceanic and Atmospheric Administration (NOAA). In October 1985, the Landsat system was commercialized and shortly afterwards all of the data holdings from GSFC and other sources were transferred to the EROS Data Center (EDC) at Sioux Falls, SD. All Landsat commercial rights became the property of Space Imaging EOSAT with exclusive sales rights to all TM data. Throughout these changes, the EDC retained primary responsibility as the Government archive of Landsat data. The Land Remote Sensing Policy Act of 1992 (Public Law 102-555) officially authorized the National Satellite Land Remote Sensing Data Archive and assigned responsibility to the Department of Interior. All Landsat data over ten years old is available from the National Archive at the EROS Data Center. In addition to its Landsat data management responsibility, the EDC investigates new methods of characterizing and studying changes on the land surface with Landsat data.

Image Data Acquisition

Initially, attention was directed towards locating TM imagery at sites throughout the Caribbean in an attempt to map the temporal signal of increased algal coverage along the pathway of Diadema mortality which began in 1982 off the eastern shore of Panama. The die-off slowly moved both north and east eventually spreading throughout the Caribbean and most of the Western Atlantic (Lessiois, et. al., 1984). We chose sites where in-situ data describing pre-Diadema and post die-off ecological conditions existed. These sites included Carysfort Reef (Key Largo, FL); Discovery Bay (Jamaica); St. Croix, (USVI); and Turneffe Reef (Belize). We also searched for imagery from Montechristi National Park (Dominican Republic) and Andros Island, (Bahamas) (Appendix I).

At first we thought it would be a relatively straightforward process to search the archive records at the EROS Data Center for available imagery, examine the thumbnail preview images for quality, and then order what we needed. Unfortunately, after an extensive amount of time searching for suitable high quality imagery of the proposed study sites, we discovered that locating and accessing the early archives of Landsat TM imagery was not a simple task. On January 1, 1984, approximately two years after launch, the Thematic Mapper was declared to be fully operational and its control was transferred from Goddard Space Flight Center to the Eros Data Center. During the first two years after launch, NASA/GSFC collected approximately one TM scene per day and the data were archived at GSFC in the "Scrounge File". The catalog of the Scrounge File and all the data files are archived at the EROS Data Center, but their specific web address is not well known, even to many of the offices within the Center. This collection of tapes in the Scrounge File is not completely indexed and is not actively included in any data sets

presently available to users. Had we not been made privy to independent information from a NASA/GSFC researcher, we would have never known of the existence of the Scrounge File (Appendix II). To date, these images are not yet catalogued and still in their original NLAPS format (Appendix IV).

Over 120 images of the prospective study sites were carefully examined. We could not locate any images of the Caribbean sites previous to the Diadema die-off in 1983 and most of the images between 1984 and 1986 were cloud covered. The northern Florida Keys, however, had a much larger image set than other locations. We found images among these that clearly showed the study site at Carysfort Reef in the years before and after the 1984 Diadema die off. Additionally, imagery of the Keys was made available from the image archives at the Jet Propulsion Laboratory (JPL) and the U.S. Geological Survey (USGS), via the Florida Marine Research Institute (FMRI). Of particular interest was an image of the Florida Keys from August 1982. This image was originally supplied to JPL by NASA/GSFC. This image was collected about a month after launch and is the earliest TM image of the Florida Keys. We were unable to locate it in any of the EDC archive catalogs, although we knew it was actually taken and should exist somewhere. Eventually, the EDC did locate the tape and we purchased a copy of the image. Finally, a time series of 22 images was assembled for the northern Florida Keys from 1982 to 1996 (Table 1, Appendix III).

Carysfort Reef lies at the northern seaward edge of the Florida Keys National Marine Sanctuary, where it is positioned at the seaward edge of the shelf. Water transparency there is usually high. We chose to focus this project on Carysfort Reef (Figure 1) because it is the largest reef in the Florida Keys and would be visible in satellite imagery. There is a long record of coral reef work at Carysfort Reef as it has been a study site in 1974. At that time, it was the richest, most diverse reef in the northern Florida Keys (Dustan, 1985b) and we reasoned that if the Florida Keys reefs were to change, it would be evident at such a site. The reef was first surveyed in 1974-5 in a study that included 21 25 meters long transects positioned from the reef flat to the deepest depth of corals (20m) on the inner reef terrace. The second census occurred in 1982-83 (Dustan 1985, Dustan and Halas 1987). This project stands as the longest, most complete study of reef coral populations in the Florida Keys. Since its inception, the site at Carysfort Reef has also been included in other reef monitoring projects, principally a series of multi-method surveys featuring photostation work by White and Porter (1985) and Porter and Meier (1992). Presently, it is one of the monitoring sites for the US EPA Coral Reef Monitoring Project (Dustan et. al., 1996)

In 1975, the shallow seaward reef zones were dominated by populations of Acropora palmata, elkhorn coral, which formed large monotypic stands so typical of Caribbean coral reefs (Goreau, 1959). A. palmata was the "signature coral" of the Florida Keys shallow water reefs. On Carysfort Reef, the highest densities of living corals occurred on the seaward edge of the shallow reef flat, the Acropora palmata zone. However, since 1975, most of the A. palmata colonies have died and the species is now being considered as a possible candidate for the Endangered Species List. Living coral coverage, estimated at 40 percent in 1975, increased to an estimated 60 percent in 1982-3 and then began to fall. The increase between 1975 and 1983 was principally due to physical breakage. Large, three-dimensional colonies were broken and lay strewn about the reef substrate. This effectively increased coral coverage while decreasing mean colony size (Dustan and Halas, 1987). By 1985, many of these coral fragments died and many regions of the former A. palmata zone were covered with algal turf communities. This began to shift the bio-optical properties of the reef towards the green spectrum with a decreased albedo.

The mass mortality of Diadema antillarum occurred in Florida during August of 1984 which further exacerbated conditions on Carysfort Reef through a dramatic reduction in grazing pressure (Hughes, 1994). Algal turf communities began to form in large areas of the shallow reef that were encrusted with highly reflective, crustose coralline algae. During this time there were more frequent outbreaks of Black Band Disease and there seemed to be an overall degradation in water quality (Dustan, 1999). High rates of mortality were documented between 1984 and 1991 (Porter and Meier, 1992). Then, in the 1996, mortality rates increased again with the outbreak of a seemingly new virulent form of White Plague Disease also termed White Pox. This disease specifically attacks A. palmata, killing large colonies within months (J. Porter, Per. communication). The present coverage by A. palmata on Carysfort Reef is estimated at 10 percent (US EPA Coral Reef Monitoring Project). Similar ecological events have occurred on many reefs throughout the Florida Keys, including Molasses Reef, Looe Key, and Sand Key. However, Carysfort is the only reef where this long-term change has been documented with quantitative line transect studies (Figure 2).

Coral reefs develop to their greatest extent in clear oligotrophic tropical seas where light intensities are high and nutrients tend to be the limiting factor. The high species diversity of coral reefs is reflected in the wide array of optical signatures emanating from the living reef substrate. While it is probable that the important optical properties of the reef are related to photosynthetic pigments fixed in the benthic organisms, there are many other organisms that add both active and passive "color" to the reef (Meyers, et. al., 1999 in press).

Alterations of many other physical and ecological parameters (wave action, water temperature, nutrient loading, herbivory, etc.) result in additional shifts in the community composition of the reef substrate, all of which alter the optical properties of the reef. Thus, the optical signal that upwells from a coral reef will be a blend of many complex spectral signatures due to the spectral complexity of the habitat, altered by photophysiology, geomorphology, and the overlying water column. As one moves away from the small scale of individual organisms, the signal becomes the product of a mixture of individual spectra. To relate an optical (remotely-sensed) signal to what is actually on the bottom therefore, one must posses knowledge of the active and passive spectral signatures of the organisms that comprise the coral reef community as well as an understanding of the effects of scale on spectral mixing.

The formidable task of remote sensing is to relate the optical signal received in space above the atmosphere to the reef substrate, living or dead. The base reef rock, calcium carbonate, is a bright white reflective substrate that rapidly becomes less reflective as it becomes covered with living organisms (Figure 3 A). Since a very large portion of the habitat is composed of living organisms, much of the upwelling reef signal is a function of the bio-optical state of the community. The upwelling spectral signatures of reef zones are, for the most part, attributable to some combinations of photosynthetic pigments that are harbored by the community members. These and other components blend together to give reef zones characteristic spectral reflectance properties that may be diagnostic for distinguishing shallow, tropical, benthic habitats. Stony and soft corals tend to be yellow-brown due to the peridinin-chlorophyll-protein complex of their zooxanthellae (Dustan, 1982) and this is clearly seen in the spectral signature of a reef zone that is dominated by stony corals (Figure 3, B, C, D). The chlorophyll of green algae and seagrasses give them their diagnostic color (Figure 3, D). Algal mat mats and turf algae tend to absorb strongly (Figure 3 A, #3). Red calcareous algae are highly reflective and reflect in a broad band between 500 and 600nm (Figure 3, C, #2).

In an effort to determine how a broadband multi-spectral TM sensor would measure the reflectance properties of individual organisms and substrates, we scaled high spectral resolution measurements of selected corals, algae, and reef substrate to the broad bands of the TM sensor. Reflectance data were supplied from two instruments. The first was a reflectance spectroradiometer (Biospherical Instruments PRR-300) and the second was a continuous grating array spectrofluorometer (Meyers et. al. in press.) that measured the spectrum continuously from 300 to 750 nm. The spectral responsivity of Landsat TM red, green, and blue sensors was used to generate conversion coefficients. These were combined with high-resolution spectra to simulate the TM broad band integrals. In order to minimize problems due to variations in illumination, we calculated blue/green and red/green ratios for comparison with the data from the TM sample subsets (Figure 4).

The data suggest that green/blue ratios may contain useful information than green/red ratio simply due to the magnitude of the differences between the ratios. Degrading the high resolution, narrow spectral bands into broad band results in a loss of information as the signals are flattened by combining wavelengths into the TM bands. The data suggest that live and dead Acropora palmata, a dominant shallow water coral, can be distinguished, but separating dead coral from rubble, or fire coral is more difficult. However, recent modeling studies on radiative transfer of these signals suggest that the blue and green wavelengths propagate through the water column and atmosphere more efficiently than longer red wavelengths (Lubin et. al, in prep).

Erdas Imagine^TM was used for all image-processing routines. Images were obtained from the EROS data center on CD-ROM. These images were in BIL (Band Interleaved) format and could be directly imported into Erdas Imagine^TM. The 1982 image obtained from the Scrounge File at EDC was in an NLAPS format and the header information could not be recovered off the tape. Images from Florida Marine Research Institute were supplies on 8mm digital tape in Erdas LAN format and converted to the Erdas IMG format. However, the images from USGS (Rictor Stumph and Ellen Rabe), supplied by Florida Marine Research Institute, had been geo-referenced and preprocessed for another project by USGS using PCI software.

The locations of reefs were gleaned from a satellite image of the former Key Largo National Marine Sanctuary that had been processed to show locations and assess the percent cover of benthic habitats. This image product was a combination of SPOT panchromatic and SPOT XS imagery. The image had been geo-referenced to standard NOAA navigational charts and a hybrid supervised-unsupervised classification applied within the Sanctuary boundaries (Jensen and Dustan, unpublished). The reef classes were shallow reef areas with less than 1-2 meters of water overlying the reef. Focusing on these shallow reef habitats minimized the effect of water column attenuation altering the spectral shape or magnitude of the reflected signal.

The pixels showing reef locations were used as areas of interest (AOIs) to subset pixels from the image series (Figure 5). Eight areas were subset including two reefs, Molasses and Carysfort, two sandy areas in shallow water (White Banks), a mangrove island (Rodriguez Key), and a region of the Key Largo mainland. This set provided a set of experimental and reference samples for trying to understand the magnitude and phenology of change. The reefs are the areas of interest. The two shallow sandy areas represent a high reflective end member that should not change very much over time. The mangrove island is an undeveloped vegetated terrestrial habitat. The Key Largo subset is an urbanized area has changed dramatically from 1982 to the present. The offshore AOI, in the axis of the Florida Current, represents the bottomless ocean, a low reflective end member that should be relatively invariant. Only subsets that were cloud free were used to compile statistics, which generated a variable number of subsets for each AOI.

We quickly realized that the image-to-image registration was imprecise. Sites were frequently 6 to 12 pixels out of registration. Careful study revealed that the images had been geo-referenced with points on land and in Florida Bay. No points over the reef line (or open ocean) had been employed so there were greater errors in the offshore waters along the "unreferenced" eastern edge of the images. Images were carefully compared and registered to the image that appeared to have the most accurate geo-registration. Each subset AOI was again carefully compared to insure image-to-image registration.

For Carysfort and Molasses Reefs, a SPOT panchromatic image was used to generate a high-precision area of interest subset (AOI). Coordinates from three independent GPS systems (differential and non-differential) were used to check the geo-referencing of the reef outline. When the AOI was overlaid on images, we discovered that the GPS coordinate fell approximately 0.3 miles to the north of the reef. Thus, while the images were co-registered, the absolute longitude and latitude were not precisely located. This required that each AOI be precisely fit to provide a "best fit" image-to image registration (Figure 6).

The pixels from each selected habitat AOI were subset and the mean brightness values (DNs) for each band (blue, green, and red) were compared between cloud free images (Table 2). Such data form the basis of mapping habitats with cluster and principle components analysis (Jensen, 1996). Tri-color plots of the DN values show that the habitats are reasonably distinct from each other, but there is between image variation within each AOI (Figure 7). The offshore waters displayed the least variation and the terrestrial urbanized environment, the most. Molasses and Carysfort Reefs are moderately variable.

When the data are examined as a time series, the two reefs exhibit high blue reflectance in 1982 which then drops over time to a minimum in. On Carysfort, values then rise slightly through 1993, then drop in 1996. Red DN values parallel the blue values. Green DN values decrease at first, then rise through 1988 and then under another similar dip through 1993 (Figure 8). It was interesting to note that the band 4 (NIR) of Rodriguez Key show a systematic decrease from 1984 to 1996, suggesting that perhaps plant biomass had decreased (Figure 9).

The temporal patterns of reef reflectance are suggestive of communities changing from a coral dominated substrate towards increasing algal coverage. The reef albedo (general reflectivity) decreases and shifts spectrally towards a greener reflectance as algal turf replaces corals. However, because some images were preprocessed differently than others, comparisons of DN values over time might be misleading. In an effort to avoid introducing radiometric errors band ratio were employed. Such normalized measures generally avoid the need for identical preprocessing, although there is a loss of analytical power. Additionally, band ratio calculations form the basis for calculating terrestrial vegetation indices and marine bio-optical photosynthetic pigment content (i.e. SeaWifs, ocean chlorophyll). Band ratio changes on Carysfort Reef shows an increase in G/B and G/R ratios peaking in 1988, then decreasing through 1993 (Figure 10). The shifts at Carysfort Reef seem to corroborate the temporal trends in DN values and are consistent with the optical signature of a substrate with and increasing algal population. Molasses Reef exhibits more variability in the G/R ratio and less in the B/G ratio.

Texture is an index of similarity of a connected pixel set. The greater the difference between connected pixels, the great the texture or spatial heterogeneity (Jensen, 1996, Hsu and Burright, 1980). The spatial heterogeneity of the subset AOIs was calculated using a 3x3-convolution kernel that was passed over the image subsets. The product depicts the degree of spatial heterogeneity, or dissimilarity of pixel distribution. Such a measure might be thought of as a measure of habitat patchiness. Since reefs exhibit distinct patterns of zonation, a measure of spatial patchiness might reflect the distribution of communities and some indication of the evenness of their distribution. An examination of the data for Rodriguez Key, a small, undeveloped mangrove island, yields values ranging from 1 to 3 with the blue band being the most variable. The temporal patterns of texture increase slightly from 1984 to 1988. From 1988 to 1996, the patterns are variable ( Figure 11). Molasses and Carysfort Reefs exhibits higher spatial texture than Rodriguez Key. Values ranging from 1.5 to almost 8. Texture on Carysfort Reef does not appear to have clear trends, although the blue band decreases from 1982 to1988, then rises and falls. Again, Molasses Reef shows a more pronounced pattern with peaks in 1985, 1991, and 1996.

While it is speculative to attach a biological meaning to texture analysis in these cases, the heterogeneity could be the result of habitat patchiness or annual phenology. Clear trends in decreasing patchiness would be consistent with the observed loss of corals and subsequent overgrowth of the substrate by algae. The loss of coral cover would be recorded as increased texture. The shallow water corals tend to be distributed in zones and patches within areas that may actually be super colonies produced through breakage and subsequent recovery. If these patches were to die and then, in turn, be replaced with algae- the spatial heterogeneity of the reef substrate would increase and then decrease.

Texture analysis can also be applied to the temporal domain to show change over time (Dobson and Jensen, 1999). Pixel-to-pixel comparisons are employed to examine change at the location of each pixel of the AOIs rather than the mean value of the AOI. In these comparisons, the AOIs must be precisely registered so as not to introduce error into the temporal comparison. Temporal texture computes the standard deviation over time at each x, y pixel location and produces an image in which a high pixel value reflects high variability over time. Temporal texture, therefore, can be employed to ask if some locations of the reef are more dynamic than others.

Analysis of Rodriguez Key (Figure 12a), reveals that the terrestrial vegetation has been relatively even while the shallow water zones surrounding the island have changed. The mainland of Key Largo (Figure 12b).

Temporal texture on Molasses and Carysfort Reefs shows that the seaward shallow portions of these reefs have the greatest amount of observable change. The highest values occur on the seaward edge of the shallow reefs, the Acropora palmata zone (Figures 13), (Figure 14). These sections of the reef have undergone extensive change. Since 1975 a large fraction of the A. palmata colonies have died and the substrate has become overgrown with algal turf. As mentioned previously, living coverage on Carysfort Reef was estimated at 40 percent in 1975. Coverage increased to an estimated 60 percent in 1982-3 (Dustan and Halas, 1987). Between 1983 and the present, coverage decreased to an estimated 10 percent (Coral Reef Monitoring Project).

The changes we have observed in the assembled imagery time series are in agreement with what is known about change in the Florida Keys. The process of urbanization has changed the mainland and temporal texture has accurately depicted this. Rodriguez Key is a natural mangrove island, which has been relatively undisturbed. The observed changes on Carysfort Reef suggest that the greatest rates of change occurred in reef zones where there is documented near-catastrophic change. The once dominant corals have died and the substrates are overgrown by turf algae. Measures of spectral reflectance suggest that the reef has become greener and albedo has decreased. The record, however, is variable and more images are needed to generate a more definitive time series. Temporal texture raises the possibility that the degree of variability may be diagnostic without added spectral information. With the addition of more images, it may be possible to resolve the issue with more certainty.

The fossil record suggests that reefs systems have been relatively stable over the long term, and now appear to be subject to an environment altered by anthropogenic effects on local regional, and global scales (Jackson, 1992). The changes that are occurring on reefs in the Florida Keys reflect other reefs throughout the Caribbean and Western Atlantic (Lang, et al 1998). However, most other reef systems do not appear to be as severely impacted as the Florida Keys, but some are clearly in worse condition (Tomascik, T. and F. Sander, 1987). In addition, reefs in the Pacific and Indian Ocean are coming under increasing amounts of stress (see Wilkinson, 1998). As these reefs change from coral to algal dominated communities, the rates biogeochemical processes of the reefs will become altered. Reef productivity may increase or decrease. The role reef carbon storage is still open to debate, but it is safe to say that, over the long term, healthy reefs tend to store calcium carbonate and export small amounts of carbon (Kawahata, et. al. 1997). Algal reefs would probably store less calcium carbonate and, perhaps, export more carbon. Since corals form the reef framework overall processes of reef growth would slow and the three dimensionally of the habitat would become reduced. As the framework of the habitat becomes degraded, other elements of the community will change and in turn, impact additional biogeochemical aspects of the reef system (Rogers, 1985). While it may not be possible to directly detect shifts in species, it does appear possible to detect community-scale change using satellite imagery from the Thematic Mapper. Reef biologists will always demand higher spatial and spectral resolution, but change detection at the scale of this study may begin to provide analyses that are useful in the management of marine sanctuaries and marine protected areas. Furthermore, the possibility exists of linking broad scale change detection techniques to critical regions that have been identified by map-based information systems (Bryant, et. al. 1998).

This work was accomplished with the help of numerous persons. Imagery was graciously provided by Rick Stumpf, USGS, and Ellen Rabe, FMRI. Lock Stuart, NASA/GSFC provided the "keys" to the Scrounge File at EDC and Tom Logan, NASA/JPL helped to locate early Landsat TM imagery. Eric Dobson contributed the image processing calculations of temporal texture. Thanks to Michael Brill, Vladimir Kosmynin, James Leard, and George Nelson for advice and technical expertise.