Coral Reefs

, Volume 26, Issue 4, pp 819–829

Bathymetry, water optical properties, and benthic classification of coral reefs using hyperspectral remote sensing imagery

Authors

    • Department of Zoology and Center for Marine BiologyUniversity of New Hampshire
  • C. D. Mobley
    • Sequoia Scientific, Inc
Report

DOI: 10.1007/s00338-007-0271-5

Cite this article as:
Lesser, M.P. & Mobley, C.D. Coral Reefs (2007) 26: 819. doi:10.1007/s00338-007-0271-5

Abstract

The complexity and heterogeneity of shallow coastal waters over small spatial scales provides a challenging environment for mapping and monitoring benthic habitats using remote sensing imagery. Additionally, changes in coral reef community structure are occurring on unprecedented temporal scales that require large-scale synoptic coverage and monitoring of coral reefs. A variety of sensors and analyses have been employed for monitoring coral reefs: this study applied a spectrum-matching and look-up-table methodology to the analysis of hyperspectral imagery of a shallow coral reef in the Bahamas. In unconstrained retrievals the retrieved bathymetry was on average within 5% of that measured acoustically, and 92% of pixels had retrieved depths within 25% of the acoustic depth. Retrieved absorption coefficients had less than 20% errors observed at blue wavelengths. The reef scale benthic classification derived by analysis of the imagery was consistent with the percent cover of specific coral reef habitat classes obtained by conventional line transects over the reef, and the inversions were robust as the results were similar when the benthic classification retrieval was constrained by measurements of bathymetry or water column optical properties. These results support the use of calibrated hyperspectral imagery for the rapid determination of bathymetry, water optical properties, and the classification of important habitat classes common to coral reefs.

Keywords

Coral reefsRemote sensingOptical propertiesHyperspectralBenthic classification

Introduction

One of the most important aspects of coral reef ecology is to monitor changes in the distribution and abundance of coral reef organisms. Our understanding of reef ecology, however, is still hampered by the inability to consistently map and monitor large expanses of reef area accurately. A significant amount of anthropogenically induced change has occurred on reefs in recent years, including coral bleaching, and phase transitions in coral reef community structure from coral-dominated to algal-dominated reefs mediated by the loss of important herbivores and trophic cascades (Hoegh-Guldberg 1999; Lesser 2004). Both bleaching and phase transitions in the reef community structure have resulted in significant coral mortality on reefs and general degradation of reefs worldwide (Hoegh-Guldberg 1999; Lesser 2004). Our ability to catalog coral reefs globally, and then assess the temporal and spatial scale of these changes has been hampered by a limited number of appropriate tools (Hochberg and Atkinson 2003).

One way to assess the coverage of coral reefs over large spatial and temporal scales is to use remote sensing imagery (Mumby et al. 1998, 2004a, b; Hedley and Mumby 2002; Purkis 2005). Remote sensing, however, requires the collection of reflectance spectra to categorize end members, the measurement of the optical properties of the water column, and radiative transfer modeling. In particular, the acquisition of reflectance signatures of the benthic community end members can be used to establish “spectral libraries”, which together with remote sensing imagery can be used in reef bottom classification analyses (Hochberg and Atkinson 2000; Hochberg et al. 2003, 2004). From a practical standpoint the testing of new optical approaches that includes remote sensing on shallow water ecosystems is best conducted on coral reefs. Coral reefs are generally found in optically clear waters (either Case 1 or 2; Mobley et al. 2004) and provide an excellent test of our abilities to extract benthic properties of reefs using remote sensing imagery. Progress in this area was initially made using multispectral data from coral reef environments. Mumby et al. (1997, 1998) compared multispectral imagery from the compact airborne spectrographic imager (CASI, used in a multispectral configuration) to satellite (Landsat, SPOT), and aerial photography. Their study showed that CASI consistently outperformed satellite sensors and aerial photography in classifying bottom features (e.g., live corals, sand, seagrass). An analysis of multispectral data from IKONOS satellite imagery (Andréfouët et al. 2003), and a radiative transfer approach with airborne multispectral platforms (Isoun et al. 2003) both provided levels of accuracy in benthic classifications that were similar to that obtained by using the CASI sensor.

Emerging techniques such as hyperspectral remote sensing from airborne and satellite platforms hold the promise of providing more detailed information than multispectral imagery. Hyperspectral data will be required for mapping coral reefs at meter-scale resolution, and to allow users to consistently discriminate mixtures of functional classes such as macroalgae and corals, as the mixing of similar spectra of these end members complicates habitat classification. Hyperspectral sensors can provide large-area coverage and imagery with sufficient spectral information to obtain water optical properties, monitor coastal interactions, and classify benthic communities. Hochberg and Atkinson (2000) and Holden and LeDrew (1999) used hyperspectral imagery and spectral reflectance libraries with a derivative analysis of those reflectances and linear discriminate function, or principal component analysis, to discriminate between coral, algae, and sand habitats on coral reefs.

To expand the number of end member or habitat classifications accurately measured in hyperspectral imagery, novel analytical approaches are required to take advantage of the additional information contained within hyperspectral imagery. In particular, the development of algorithms capable of consistently extracting accurate bottom classifications from remote sensing reflectance (Rrs, the ratio of upwelling water leaving radiance to downwelling irradiance) is required. Addressing substrate heterogeneity has been attempted using derivative analysis and linear un-mixing techniques with varying success because of the non-linear nature of benthic reflectances, and their mixing, as they propagate to the surface through the water column (Hedley and Mumby 2003; Hedley et al. 2004). Another method that addresses the problem of spectral mixing is to blend different algorithms in a “fuzzy” logic classification scheme. Fuzzy logic classification allows for multiple end members to be classified in an individual pixel. This type of classification scheme would reduce errors associated with one-pixel one- end-member algorithms and better reflect the heterogeneous nature of coral reef habitats on small spatial scales. Fuzzy classification schemes have been used with remote sensing imagery for assessing both primary productivity and benthic classifications (Andréfouët and Roux 1998; Andréfouët et al. 2000; Moore et al. 2001). Another technique recently described by Louchard et al. (2003) and Mobley et al. (2005) can extract benthic classifications, bathymetry, and inherent optical properties (IOPs: in particular, the absorption, scattering, and backscatter coefficients) from hyperspectral imagery. They characterized benthic habitats into defined classes by comparing the Rrs of individual pixels in hyperspectral images to Rrs from a library of simulated Rrs spectra created using radiative transfer modeling and measurements of benthic reflectances for pure end members and mixtures representing specific habitat classes on coral reefs. This spectrum-matching look-up table (LUT) classification scheme has been proposed as an alternative approach to using one-dimensional radiative transfer models (Lyzenga 1981; Maritorena et al. 1994) that calculate the attenuation of radiance at individual wavelengths through the water column and assume that the IOPs are homogeneous and the bottom is a perfect Lambertian reflector.

Previous utilization of the LUT protocol was restricted to simple benthic classifications in shallow waters with end members of simple spectral composition (Louchard et al. 2003; Mobley et al. 2005). In the present study, the LUT approach was evaluated for the extraction of bathymetry, water column IOPs, and benthic classification using hyperspectral Rrs imagery collected from a shallow reef (4–12 m) near Lee Stocking Island (LSI), Bahamas. This coral reef habitat provides a more rigorous test of the LUT protocol to extract benthic classifications from hyperspectral Rrs data that is significantly affected by the spectral mixing of several end members in the community.

Materials and methods

Field sites and sample collection

Hyperspectral imagery and all ground truth data described below were collected in May 2000 during the Coastal Benthic Optical Properties program at Horseshoe Reef, LSI, Bahamas (23°46.5′N, 76°05.5′W).

Hyperspectral imagery

The Ocean Portable Hyperspectral Imager for Low-Light Spectroscopy (Ocean PHILLS) was deployed on an Antonov AN-2 aircraft around LSI in May 2000. Imagery was obtained for specified flight lines in the morning (09:00–10:00 h) to achieve a solar zenith angle of 40–55°. Flight lines were run at an azimuthal angle of 83° to be aligned with the solar azimuth. The PHILLS spectrometer is a pushbroom-scanning instrument (Davis et al. 1999) that measures 512 spectral channels at 1,024 spatial across-track samples. The spectral channels, however, are usually binned in increments of four on the charge-coupled device chip to yield 128 channels of output, each approximately 4.5 nm wide. Typically the aircraft flew at a 2,592 m altitude and 100 knots, resulting in a spatial resolution on the ground of approximately 1.3 m. A spectral calibration of the PHILLS was performed in the laboratory by imaging several different gas lamps (oxygen, mercury, argon, and helium). By pairing up measured emission lines with known lines, a relationship was derived between channel number and wavelength of the center of the channel. A pre-deployment radiometric calibration was performed using a 40 in Spectralon-coated integrating sphere containing ten halogen lamps (Labsphere, Inc., North Sutton, NH, USA). Measurements were made of this diffuse source with a number of various lamps turned on. The intensity of the sphere at each lamp levels is known from independent calibration. Because the sphere source is red rich, a blue filter was placed in front of the PHILLS lens to make the source more spectrally flat. The data were atmospherically corrected to produce remote sensing reflectance using the TAFKAA atmospheric correction algorithm (Gao et al. 2000; Montes et al. 2001). In the process, atmospheric absorption lines were used to make adjustments in the spectral calibration.

Some pixels in the imagery appeared to be contaminated by whitecaps or sunglint. Any pixel with Rrs (800 nm) > 0.007 sr−1 was flagged as a whitecap and not processed. The remaining atmospherically corrected Rrs spectra were systematically positive by approximately 0.003 sr−1 at wavelengths greater than 800 nm. The water-leaving radiance should be 0 at these wavelengths because of high absorption by the water itself; these very clear waters have low particle concentrations that could result in non-zero spectra. The non-zero Rrs in the near-IR indicates an undercorrection for atmospheric path radiance, which is consistent with the assumed low wind speed of 2 m s−1 that was used in TAFKAA. Therefore, 0.003 sr−1 was subtracted from each Rrs spectrum before processing by the LUT algorithms.

Figure 1 shows a red-green-blue image of the study area where the flight lines were flown on and around LSI. The yellow box in Fig. 1 is the 250 pixel × 250 pixel (325 m × 325 m) region of interest containing Horseshoe Reef for this study. The area within the yellow polygon shows the subset of the PHILLS study area analyzed within Horseshoe Reef. Visually it is easy to discern that Horseshoe Reef is a shallow reef area surrounded by shallower hardpan shoreward and sand, seagrass beds, and deeper open water seaward of the reef. For the habitat classification analysis, only the portion of the reef area within the yellow polygon was used for comparison with the transect study of community structure conducted on the reef site as described below.
https://static-content.springer.com/image/art%3A10.1007%2Fs00338-007-0271-5/MediaObjects/338_2007_271_Fig1_HTML.jpg
Fig. 1

Red, Green, Blue (RGB) image of Horseshoe Reef, Lee Stocking Island, Bahamas, showing the area (yellow box) of analysis, and the image of the area of analysis within the yellow polygon. Only pixels within the polygon were used for comparison with the benthic survey data taken on the transect lines along the reef

Reflectance measurements

A diver-operated spectrometer (Mazel 1997, www.nightsea.com/divespec.htm) was used to measure reflectance spectra of all functional end members (corals, macrophytes, seagrasses, and sand) in situ. The unit contains a Model S2000 Spectrometer (Ocean Optics Inc.) controlled by a Tattletale Model 8 data logger (Onset Computer Inc.). Light reaches the spectrometer via a 1.5-mm liquid light guide cable with a nominal 5 nm spectral resolution. The spectrometer is fitted with a grating/detector combination that records 2,048 spectral bins covering the wavelength range from approximately 300 to 850 nm. A standardized full-spectrum light source is provided by an array of red, blue, and white light-emitting diodes directed through a diffuser to provide even illumination at approximately normal incidence. The probe head excludes ambient light and fixes the light guide at a 45°angle to the measurement surface. The operator records a reference spectrum by measuring the light reflected from a Spectralon reflectance standard. Reflectance is computed as the ratio of the light reflected from the sample to the light reflected from the standard. This measurement is equivalent to the irradiance reflectance (Rb) if the material is a lambertian reflector, which is assumed to be the case. The reflectance spectra were interpolated to the LUT database wavelengths, namely 400–750 nm at 5 nm resolution, for use in the HydroLight modeling described below.

Measurements of inherent optical properties

Water column IOP measurements, specifically the spectral absorption a (λ) and beam attenuation c (λ) at wavelengths 412, 442, 488, 510, 532, 589, 620, 676, 715 nm, were made with an ac-9 (WET Labs, Inc.) at Horseshoe Reef and at several other locations around LSI at the same time of imagery acquisition.

Hydrographic survey

A hydrographic survey was conducted by the Naval Research Laboratory in 1998 from a small (7.62 m) boat using a temporary global positioning system (GPS)/Transducer mount that was retrofitted to the boat and held the GPS antenna directly over the echo sounder-transducer. Bathymetric measurements were recorded at a repetition rate of 0.7 Hz using a Suzuki ES2025 echo sounder, and each depth was recorded along with its time and the latitude and longitude as determined by the Wide Area Augmentation System GPS. The acoustic data were corrected to mean sea level (MSL) to account for tide differences, and the average survey speed of the vessel was 7 knots. A survey of the North (windward) side of LSI was conducted and all the data were collected with horizontal reference to the WGS-84 ellipsoid, in Universal Transverse Mercator coordinates, zone 18, with the central meridian at 75° West. All soundings were corrected to MSL in real time by utilizing the Z axis output of the GPS receiver.

All the sounding/positional data were reviewed and edited manually to remove erroneous entries. There were very few data points deleted as a result of the editing process. Differential kinematic corrections were maintained throughout all surveys. Some soundings were deleted due to transducer broaching in heavy seas. All the positional data used were corrected at the time of collection and it is expected that the positional accuracy is within the GPS receiver manufacturer’s specification of 20 cm circular error probable. All soundings are in meters below MSL.

Benthic surveys

The habitat assessment at Horseshoe Reef was characterized by employing three 10 m transect lines parallel to the seaward contour of the reef at a depth of 6–8 m at three sites on Horseshoe Reef (Fig. 1, within the polygon). The benthic habitat type was recorded using 15 randomly placed 0.25 m2 photographic (35 mm) quadrats along each 10 m transect line using a Nikonos V with a 15 mm lens and synchronized strobes using a quadropod with scale (2 cm). Individual photographs were digitized into JPEG images and the projected surface area of the dominant functional groups was analyzed using NIH Image software (http://www.rsb.info.nih.gov/nih-image/). For each photograph the entire quadrat was analyzed for the percent projected surface area of hard corals, macrophytes, seagrasses, turf algae, sand, and bare consolidated carbonate substrate and expressed as percent cover for each end member. These benthic survey data were then used to define different benthic habitat classifications for comparison with the LUT bottom retrievals within the Horseshoe Reef area outlined in the yellow polygon (Fig. 1).

Look-up table generation and classification protocol

A database of remote sensing reflectance spectra Rrs from 400 to 750 nm by 5 nm bandwidths, corresponding to various water depths, bottom reflectance spectra, and water column IOPs was assembled using a special version of the HydroLight radiative transfer numerical model (Mobley 1994). HydroLight includes all orders of multiple scattering, and its numerical algorithms are tuned for accuracy rather than for minimization of computation time. It thus provides solutions of the unpolarized radiative transfer equation that are accurate to approximately one percent for the given input. Each HydroLight-generated Rrs spectrum in the database is tagged by indices that identify the bottom depth, bottom reflectance spectrum, and IOP spectra that were used as input to the HydroLight run. At a minimum, this database needs to contain Rrs spectra generated for environmental conditions similar to those occurring in nature at the time and location where the image was acquired.

For the present analysis, the Rrs database was created for a sun zenith angle of 50°, clear skies, a wind speed of 5 m s−1, and a nadir viewing direction. These conditions are appropriate for the image being processed. Seven sets of water absorption a and scattering b spectra spanning the range of IOP spectra measured at various times and locations around LSI were used for database generation. After removal of absorption by pure water, the absorption spectra showed the exponential dependence on wavelength that is characteristic of water containing dissolved organic matter. The absorption spectra were therefore extrapolated to 400 nm using an exponential function whose slope was determined by the measured values at 412 and 442. Absorption was assumed to be 0 at wavelengths longer than 715 nm. The beam attenuation was extrapolated to 400 and 750 using a power law whose exponent was determined from the measured values. The extrapolated a (λ) and c (λ) were then interpolated to the LUT wavelengths, and the scattering coefficient b (λ) was obtained from c (λ) − a (λ). For each set of a and b spectra, backscatter spectra bb corresponding to particle backscatter fractions of 0.01, 0.02, 0.03, and 0.04 were constructed from the scattering spectrum b. Thus the database has Rrs spectra corresponding to 28 different combinations of a, b, and bb. In addition to the measured end member reflectance spectra for various bottom types, mixtures of sand and seagrass reflectances were defined for combinations of 10% sand and 90% grass through 90% sand and 10% grass. Mixtures of pavement and turf algae, pavement and sargassum, and pavement and corals were similarly defined. Several additional combinations of coral, sand, and algae then gave a total of 118 bottom reflectance spectra. Bottom depths covered 0.25–15.0 m by intervals of 0.25, 15.0–25.0 m by 0.5 m, plus depths of 0.01, 30.0, 50.0, and infinity, for a total of 84 possible depths that can be retrieved. The 28 IOP sets, 84 depths, and 118 bottom reflectance spectra gave a database with almost 275,000 Rrs spectra.

When analyzing an image, the Rrs spectrum for a particular pixel in the image is compared with each spectrum in the database and the closest match to the image spectrum is found by a Euclidean least-squares minimization over wavelength. It should be noted that the LUT spectrum-matching makes full use of both the magnitude and spectral shape information available in radiometrically calibrated and atmospherically corrected spectra. The details of the spectrum matching and LUT methodology are given in Mobley et al. (2005).

Results

The image area seen in the yellow polygon (Fig. 1) was processed using the LUT methodology described in Mobley et al. (2005) and the Rrs database created with the IOPs and bottom reflectances described above. In an unconstrained LUT inversion, nothing is presumed known about the environment and the bathymetry, water column IOPs, and bottom reflectance are simultaneously retrieved. In a bathymetry-constrained inversion, the depth is known at each image pixel from the acoustic bathymetry interpolated to the pixels. The LUT spectrum-matching then searches only those Rrs spectra corresponding to the database depth closest to the known depth at each pixel, and only the IOPs and bottom reflectance are retrieved. In an IOP-constrained inversion, some or all of the IOPs are presumed known from ancillary measurements or bio-optical models. LUT then retrieves only the depth, bottom reflectance, and any unspecified IOPs.

Figure 2 a shows the average Rb spectra computed from multiple measurements for various bottom types (with 8–10 samples for each bottom type). The figure also shows two spectra corresponding to mixtures of a sand spectrum and a seagrass spectrum, weighted as 90% sand and 10% grass, and 80% sand and 20% grass. These two mixed sand-grass spectra are similar in magnitude and spectral shape to the average hardpan spectrum. Figure 2b shows the individual coral, turf algae, seagrass, and macrophyte spectra used in computing the averages of Fig. 2a. On an average, turf algae have a somewhat greater reflectance than grass, which has greater reflectance than macrophytes. However, there is an overlap in the individual spectra of turf algae, grass, and macrophytes, especially at blue and green wavelengths where the water is most transparent and the bottom reflectance has the greatest influence on Rrs. Likewise, the spectra for darker sediments such as hardpan and pavement can be similar to the spectrum of a brighter sand bottom with a sparse cover of dark sea grass or other vegetation. Thus, although the average spectra for different bottom types are optically distinct, it is not certain how well biologically different bottom types such as grass, turf algae, and macrophytes, or hardpan and sand–grass, can be optically distinguished in practice.
https://static-content.springer.com/image/art%3A10.1007%2Fs00338-007-0271-5/MediaObjects/338_2007_271_Fig2_HTML.gif
Fig. 2

a Average bottom reflectances (Rb) for various bottom types. b Individually measured reflectances (Rb) of typical bottom types for various functional groups of benthic biota used to calculate averages presented in Fig. 2a

The HydroLight computation of Rrs as described above used 118 Rb spectra, which include individual Rb spectra such as those presented in Fig. 2b, the average spectra of Fig. 2a, and mixtures such as sand–grass, pavement–turf, and coral–sand–algae in various combinations as discussed above. The retrievals for these 118 spectra were grouped into categories that are optically distinct and which roughly correspond to the compositional categories as measured in the line transects. The LUT retrieval categories are thus bare sand, darker bare sediment or sand with a sparse cover (<20%) of grass, sediment with various mixtures of grass, turf, or macrophytes, pure coral, and sediment with various mixtures of coral and algae.

Figure 3 shows a depth profile of a (λ), b (λ), and c (λ) coefficients at 488 nm as measured or determined using an ac-9 within the study area showing that the shallow waters are well mixed within the study area as expected because of strong tidal currents in this area. Measurements of IOPs using ac-9 measurements at a study site within 1 km, were similarly constant with depth during the acquisition of imagery. The optically clear waters at Horseshoe Reef typically have chlorophyll concentrations less than 0.2 mg m−3. The observed absorption coefficients were often much higher at blue wavelengths than would be expected for Case 1 waters with these low chlorophyll concentrations. The additional absorption is due to colored dissolved organic material (CDOM) derived from seagrass, corals, and other benthic biota (Boss and Zaneveld 2003). The high-CDOM waters in the shallows to the west and south of Horseshoe Reef are tidally mixed with the deep Case 1 waters to the north and east, giving a strong tidal signal in the absorption coefficient. These waters are therefore considered Case 2 because of the poor correlation between absorption and phytoplankton pigment concentration. The scattering coefficient, however, is closer to what is expected for Case 1 waters (Mobley et al. 2004, 2005).
https://static-content.springer.com/image/art%3A10.1007%2Fs00338-007-0271-5/MediaObjects/338_2007_271_Fig3_HTML.gif
Fig. 3

Depth profile of absorption [a (λ)], scattering [b (λ)], and beam attenuation [c (λ)] at 488 nm with the pure water contribution removed from ac-9 measurements

For the bathymetry surveys, the depths at each acoustic ping location were linearly interpolated in space to generate a depth at each image pixel. Figure 4 shows the interpolated acoustic bathymetry for the yellow rectangle area shown in Fig. 1, presented in 1 m depth bins from 4 to 12 m. The black dots in the figure show the locations of the acoustic pings, which are typically spaced a few meters apart along track and about 10 m apart cross-track. The acoustic data are used for validation of the unconstrained LUT-retrieved bathymetry at image pixels for which there is an acoustic ping. The interpolated acoustic data are used to constrain the depth at each pixel to a known value in depth-constrained retrievals.
https://static-content.springer.com/image/art%3A10.1007%2Fs00338-007-0271-5/MediaObjects/338_2007_271_Fig4_HTML.gif
Fig. 4

Acoustic bathymetry of the area of Fig. 1 constructed by interpolation between the acoustic ping locations (black dots), color coded in 1 m bins between 4 and 12 m. All soundings are in meters below mean sea level

Figure 5 compares the acoustic and unconstrained LUT-retrieved depths for the 2,706 image pixels for which an acoustic depth was measured (the black dots shown in Fig. 4). On average the LUT-retrieved depths are 4.5% or 0.37 m too shallow, with a standard deviation of 1.16 m. Ninety-two percent of the LUT depth retrievals are within ±25% of the acoustic depth, and 66% are within ±1 m of the acoustic depth. The outlying points in this figure, which indicate inaccurate retrievals, often come from pixels where the acoustic bathymetry shows a sharp gradient in the bottom depth. Georectification for this imagery was performed by manual image warping pegged to the location of LSI because the on-board aircraft navigation unit was not working. This becomes increasingly inaccurate as the distance from LSI increases (toward the upper right of Fig. 1). Detailed analysis of other images taken during the same flights indicates errors of several meters or more in the computed horizontal positions of individual pixels (Mobley et al. 2005). Pixel location errors mean that the LUT depth retrieval for a particular pixel is being incorrectly compared with an acoustic ping location that may be several meters away, and which may have a depth several meters different than the actual depth at the image pixel. Thus an unknown amount of the LUT depth error is due to an improper association of acoustic ping locations with image pixels, rather than to bad retrievals by the LUT algorithm per se. However, it is not possible to separate these two sources of error in the imagery used for these analyses.
https://static-content.springer.com/image/art%3A10.1007%2Fs00338-007-0271-5/MediaObjects/338_2007_271_Fig5_HTML.gif
Fig. 5

Look-up-table (LUT)-retrieved depths vs. acoustic depths for the 2,706 pixels in Fig. 1 for which an acoustic depth is available (the black dots)

For evaluation of the LUT thematic bottom classification it would be desirable to compare the LUT retrievals and ground truth pixel by pixel and to compute an error matrix of producer and user accuracies (Green et al. 2000). However, the line transect data did not allow such a detailed comparison. The best that can be done with the available benthic survey data is to compare the measured percent coverages of habitat types along Horseshoe Reef transect lines with the LUT percent coverages within the corresponding area delineated by the yellow polygon in Fig. 1. The results of these comparisons are shown in the first data line of Table 1. The notation in column 1 indicates that all seven sets of absorption and scattering coefficients, each with four particle backscatter fractions Bp, was used and that the depths were unconstrained. This is a completely unconstrained retrieval and serves as the baseline for various constrained retrievals. The measured percent coverages of the different habitat classes are shown in the last line of the table. Although the optical bottom classification categories do not correspond exactly to the biological categories, the roughly equivalent benthic classifications are in good agreement given the uncertainties of the comparison. The last column of Table 1 shows the previously stated errors in the depth retrieval.
Table 1

Results obtained for various constraints on the look-up-table (LUT) retrieval

Retrieval database

Bare sand

Dark sediment or sand and sparse grass

Sediment, grass, turf, or macrophytes

Pure coral

Sediment, coral and algae

Depth errors: % error/rms err/% in ±1 m/% in ±25%

7 (a, b) × 4 Bp; unconstrained depths

3.3

8.5

70.1

10.3

7.8

−4.5/1.16/66/92

4 (a, b) × 4 Bp; unconstrained depths

3.3

8.5

69.9

10.4

7.9

−4.6/1.20/65/90

4 (a, b), Bp = 0.02; unconstrained depths

3.3

9.1

67.3

12.0

9.0

−5.0/1.24/65/90

1 (a, b), Bp = 0.02; unconstrained depths

3.1

9.4

49.4

12.1

26.0

−1.3/1.39/61/86

7 (a, b) × 4 Bp; constrained depths

3.2

6.8

70.8

15.4

3.8

NA

4(a, b) × 4 Bp; constrained depths

3.2

6.8

70.7

15.5

3.8

NA

4(a, b), Bp = 0.02; constrained depths

2.8

5.9

68.4

16.0

6.9

NA

1 (a, b), Bp = 0.02; constrained depths

2.9

7.7

63.6

16.7

9.2

NA

Measured (mean ± SE)

2.5 ± 0.71

14.3 ± 1.12

69.8 ± 4.56

13.3 ± 0.69

NA

NA

All numbers for benthic classification are percent coverage for the reef area within the polygon in Fig. 1. The LUT values are computed using only the pixels within the polygon seen in Fig. 1; the measured values are from the line transects. The depth errors are for the 2,706 pixels of Fig. 1 for which an acoustic depth was available

NA not available, rms err root mean square error

Although the IOP data are not adequate for a rigorous LUT validation, this comparison shows that the LUT-retrieved IOPs are consistent with the available data from Horseshoe Reef and nearby locations. The backscatter coefficient for the area of analysis corresponds to a particle backscatter (b) fraction of 0.02. Measurements of the backscatter coefficient at 488 nm at Horseshoe Reef and nearby North Perry Reef (but not at the exact same time as the overflight) gave bb (488) values of ≈0.0028 m−1. Assuming that this value can be compared with any of the b values for Horseshoe Reef, the particle backscatter fraction would be in the 0.02–0.03 range at 488 nm, which is consistent with the retrieved value of 0.02.

Next it was investigated how robust the unconstrained retrievals of habitat classifications (Table 1, Fig. 6a) were if the IOPs or bottom depth was constrained (Table 1, Fig. 6b). The lowest and highest magnitude IOP spectra were first excluded from the LUT database, leaving only four sets of a and b, each with the four particle backscatter fractions. These IOPs bracket what are expected for these waters, based on the available IOP measurements. Then the backscatter fraction is further constrained to be 0.02, and finally only the one set of a, b, and bb spectra are allowed. This forces each pixel to have exactly the same IOPs. Such a strong constraint on the IOPs may be unreasonable because there may be gradients in the IOPs between the shallower (perhaps more benthic CDOM or more re-suspended sediment) and deeper areas of the image. The IOP results for the various constrained-IOP retrievals are shown in data lines 2–4 of Table 1. The first two constrained retrievals give almost the same results as the unconstrained retrieval. However, constraining the IOPs to only one possible set reduces the percent of sediment + grass, turf algae, or macrophytes by about 20% and increases the percent of sediment + coral and algae by the same amount (Table 1, Fig. 6a, b). The depth retrievals are degraded slightly when the IOPs are constrained. In these clear shallow waters, the bottom depth is often the primary factor determining the magnitude of the remote sensing reflectance. Thus LUT tends to retrieve a bad IOP before it retrieves a bad depth (Mobley et al. 2004). Constraining the IOPs thus removes LUT’s ability to adjust the IOPs in order to retain the depth retrieval, and the depth errors sometimes increase as the IOPs are constrained.
https://static-content.springer.com/image/art%3A10.1007%2Fs00338-007-0271-5/MediaObjects/338_2007_271_Fig6_HTML.gif
Fig. 6

a Benthic classification using the look-up-table (LUT) protocol and an unconstrained retrieval for the area shown in Fig. 1. The percent coverage is computed for the Horseshoe Reef area within the black polygon. b Benthic classification using the LUT protocol and a constrained retrieval for the area shown in Fig. 1. The percent coverage is computed for the Horseshoe Reef area within the black polygon

The depths were next constrained to be equal to the interpolated acoustic depth at each pixel for the same sequence of IOP constraints. The results are shown in data lines 5–8 of Table 1. Constraining the depths causes small changes in the various benthic categories, but overall the results are not much different than the unconstrained or IOP-only constrained retrievals (Table 1, Fig. 6a, b). Constraining the depths brings the single-IOP case back into closer agreement with the other retrievals.

Using the same conditions described above for the unconstrained and constrained retrievals (Fig. 6a, b) maps of bottom classifications were then produced, limited to pixels defined as pure sand, seagrasses/turfs/macrophytes, and pure corals with two other classifications of seagrasses/turfs/macrophytes, and pure corals with any additional matching to sand or sediment. These retrievals (Fig. 7a, b) produce habitat classification maps with less “noise” then previous maps (Fig. 6a, b). Both the unconstrained and constrained retrievals show a similar percent coverage of sand and coral end members within the area of analysis while retrievals of seagrass/turf/macrophtes became more segregated between pixels with and without matches of sand or sediment.
https://static-content.springer.com/image/art%3A10.1007%2Fs00338-007-0271-5/MediaObjects/338_2007_271_Fig7_HTML.gif
Fig. 7

a Benthic classification using the look-up-table (LUT) protocol and an unconstrained retrieval for the polygon area shown in Fig. 1 but constrained to an analysis of the three principal end members (e.g., sand, seagrass/turf/macrophytes, coral) and the seagrass/turf/macrophytes, and coral end members with any match to sand or sediment. b Benthic classification using the LUT protocol and a constrained retrieval for the polygon area shown in Fig. 1 but constrained to an analysis of the three principal end members (e.g., sand, seagrass/turf/macrophytes, and coral) and the seagrass/turf/macrophytes, and coral end members with any match to sand or sediment

Discussion

The spectrum-matching and LUT methodology as described by Mobley et al. (2005), and used in this study, provided information on bathymetry, water optical properties, and benthic classification for Horseshoe Reef on LSI in the Bahamas. In particular, the LUT-retrieved percent coverage for various benthic classification categories (sediments, corals, various mixtures of sediments and seagrass, turf algae, and macrophytes) compared very well with the percent coverage determined by diver transects and supports the use of this approach on coral reefs that have less than 20% coral cover and mixed spectral bottoms as observed through most of the Caribbean. Coral reef habitats with greater percent cover of corals (e.g., many areas of the Pacific) with fewer benthic end members and consequently spectral mixing would also be very amenable to this approach.

Potential sources of error in the analysis include the fact that the LUT methodology had not been developed at the time of the 2000 field experiment, and the IOP and bottom classification data were obtained for purposes other than detailed validation of the LUT retrievals. There were also problems with the imagery obtained by this initial version of the PHILLS instrument, including imperfect atmospheric correction and failure of the on-board navigation that decreased the accuracy of the image georectification. Nevertheless, the results are reasonably accurate when compared to direct measurements and the consistency of the LUT retrievals with the available data is encouraging, even though the optical and benthic classification categories could be improved with more comprehensive ground truthing data sets. The LUT retrievals are robust for the various constrained inversions considered. This indicates that the unconstrained inversion was producing accurate results.

For many coral reefs the bathymetry may be known from nautical charts or acoustic or LIDAR surveys. When that is the case, the LUT retrieval at each pixel can be constrained by the known depth, so that only water column IOPs and bottom type need be retrieved by the spectrum-matching process. If a comprehensive (spatially and temporally) set of IOP measurements is available for the overflight, then the IOPs can be taken as known. In such cases, the LUT retrieval can be constrained by the known bathymetry or IOPs, which should improve the bottom classification. Theoretically, using the LUT protocol we should be able to “remove the water” for reefs in deeper water as well, and it is possible that more comprehensive IOP data sets for these adjacent reefs, or a larger spectral library, would improve LUT retrievals for reefs deeper than 15 m. However, the LUT methodology may be limited to maximum depths of ∼20 m (Hochberg and Atkinson 2003; Hochberg et al. 2003; Mumby et al. 2004b) in even the clearest waters, because absorption by water itself inescapably reduces the contribution of bottom reflectance to Rrs in deep waters.

For the LUT protocol, constraining the IOPs and/or bottom depth thus does not greatly affect the bottom classification retrievals because the unconstrained inversion is already giving retrievals that are consistent with the available measurements and thus can be taken as correct within the limitations of our ground truth data. In other words, constraining an already good solution does not improve the solution. The unconstrained retrievals are robust in the sense that they are not greatly affected by additional constraints on the IOPs or bathymetry. If the IOPs are constrained to be outside the range of values observed in these waters, or if the depths are constrained to be incorrect, then the retrieval quality is easily degraded. For example, if all acoustic depths are multiplied by a factor of 0.75, thus forcing the bottom to be too shallow in a depth-constrained retrieval with unconstrained IOPs, then the fraction of pure corals in the Horseshoe Reef area increases to 60%, the fraction of sediment and sediment–coral–algae mixtures decreases to less than 1%, and large areas of bare sand are reclassified as dense grass. These results are clearly incorrect, as a consequence of incorrectly constraining the depths. Although constraining the retrieval does not greatly alter results, it does greatly affect the run time because far fewer Rrs spectra must be searched when constraints limit the allowable spectra for each pixel. For the present image of 250 pixels × 250 pixels and a LUT database with ∼275,000 Rrs spectra, the unconstrained retrieval required about 14 min on a 2 GHz PC. However, the depth-constrained, single-IOP retrieval required only 5 s.

While it is now generally agreed that broad-band multispectral sensors have limitations for comprehensive benthic habitat classification of coral reefs, the current group of airborne and satellite hyperspectral sensors can also be improved in pixel size and signal-to-noise properties for assessing and monitoring the global status of coral reefs (Hochberg and Atkinson 2003). Hyperspectral data, however, do provide significantly greater resolution and accuracy in the classification of submerged benthic habitats (Karpouzli et al. 2004), and in those studies that examine reef scale classification of habitat classes (Mumby et al. 2004b; this study). This study can also provide additional data toward the development of an entirely new satellite sensor for synoptic, hyperspectral, coverage of submerged benthic habitats (Hochberg and Atkinson 2003), as well as the improvement of analytical approaches that need to be evaluated for use with hyperspectral imagery.

These results presented here indicate that the combination of hyperspectral imaging and the LUT retrieval methodology for mapping and monitoring of common benthic habitats on coral reefs can be employed for reefs as deep as 10–12 m. Even if benthic classification from imagery is less accurate than the traditional methods such as diver surveys or video taken from boats, the rapid acquisition and processing of airborne imagery allows for mapping and monitoring large areas with much less effort and expense than traditional mapping techniques. The future field experiments for the LUT protocol should be designed with quantitative pixel-by-pixel LUT validation as a primary goal.

Acknowledgments

This research was supported by grants from the Office of Naval Research-Environmental Optics Program as part of the Coastal Benthic Optical Properties (CoBOP) program to MPL and CDM, who also received additional ONR support for the development of the LUT methodology, and was also made possible with funding to MPL provided by the Coral Reef Targeted Research (CRTR) Program. The CRTR Program is a partnership between the Global Environmental Facility, the World Bank, the University of Queensland (Australia), the United States National Oceanic and Atmospheric Administration (NOAA), and approximately 40 research institutes and other third parties around the world. We thank Curt Davis, Robert Leathers, Valerie Downs, Dave Phinney, Emmanuel Boss, Ron Zaneveld, Charles Mazel, and the staff of the NOAA Caribbean Marine Research Center at Lee Stocking Island, Bahamas for their help and assistance. This research conforms to the applicable laws of both the United States and the Bahamas.

Copyright information

© Springer-Verlag 2007