Journal of Coastal Research
SI
62
86–98
West Palm Beach, Florida
Spring 2011
Benthic Classifications Using Bathymetric LIDAR Waveforms and Integration of Local Spatial Statistics and Textural Features Antoine Collin{, Bernard Long{, and Phillippe Archambault{ { INRS-ETE University of Que´bec Que´bec, Canada
[email protected]
www.cerf-jcr.org
{ ISMER University of Que´bec in Rimouski Rimouski, Canada
ABSTRACT COLLIN, A.; LONG, B., and ARCHAMBAULT, P., 2011. Benthic classifications using bathymetric LIDAR waveforms and integration of local spatial statistics and textural features. In: Pe’eri, S. and Long, B. (eds.), Applied LIDAR Techniques, Journal of Coastal Research, Special Issue No. 62, 86–98. West Palm Beach (Florida), ISSN 0749-0208. The scope of this research is to assess benthoscape discrimination by airborne light detection and ranging (LIDAR) bathymetry (ALB) on the basis of statistical parameters derived from the LIDAR waveforms, textural information, and local spatial statistics. Analysis of the underwater camera stations allowed clustering of the stations into groups on the basis of their habitat composition (b-diversity). Twelve descriptive statistics describing the shape of the bottom part of the waveform, also called 12 benthic parameters, were used for discriminating four benthic habitats. A K-means classification and a supervised method based on the Support Vector Machine (SVM) were applied to this dataset, and overall accuracies of 67.7% and 89.9% were obtained, respectively. Geostatistical analyses, using 11 textural measures, defined by the gray-level occurrence matrix (GLOM) and the gray-level co-occurrence matrix (GLCM), and three local spatial statistics were then applied to the 12 benthic parameters to enhance the SVM classification performance. The assessment of the contribution of geostatistics into benthic class segmentation was achieved by computation of separability distance. Mean (from the GLOM), mean (from the GLCM) and the local Getis-Ord statistic yielded the best rates of discrimination. These added metrics, integrated with bands related to the 12 benthic parameters, showed that the rate of correct (supervised) classification was thereby improved and increased by 5.3%. Finally, the first four principal components (PCs) (i.e., 90.41% of the 12 parameter variances, boosted by the three best geostatistics) brought out an overall accuracy of 93.3%, showing evidence for optimizing the classification processing.
ADDITIONAL INDEX WORDS:
Bathymetric LIDAR, waveforms, classification, habitat, texture, spatial statistics.
INTRODUCTION Coastal ecosystems encompass a broad range of habitat types, harbor a wealth of species, and provide many ecosystem services, including nutrient storage and cycling, filtering pollutants, forming inland freshwater systems, helping protect shorelines from erosion and storms, and providing areas for leisure, recreational activities, and tourism (Costanza et al., 1997). Recent projections estimate that, by 2025, up to 2.75 billion people will live within 60 miles of the coast, creating the world’s greatest population densities along the coastal areas (Gaffin, Engelman, and Hachadoorian, 2006). However, coastal zones face significant threats from many factors, including sea level rise, increased hurricane intensity, coastal erosion, urbanization, losses of coastal wetlands and biodiversity, and marine pollution (UNEP, 2006). The vulnerability of littoral zones should thus be evaluated with a multiscale spatial approach. The development of remote sensing technologies has enabled monitoring of spatiotemporal patterns in ecological DOI:10.2112/SI_62_9 received and accepted in revision 13 September 2010. E Coastal Education & Research Foundation 2011
systems at various scales. Sensor and imagery improvements have provided meaningful insights in various landscapes, including agriculture, forest, urban, and ocean surfaces (Evans et al., 2005; Lucas et al., 2007; Ruiliang et al., 2008; Salovaara et al., 2005). However, for marine and coastal systems, patchy field and remote observational data and a lack of regional data have historically prevented quantification of patterns and processes at a landscape scale (1–100 km). The lack of efficient seafloor mapping constitutes not only a serious obstacle to understanding the structure and dynamics of assemblages of species at various scales but also hinders efforts to properly manage marine and coastal habitats, particularly with respect to fisheries resources, for which most stocks are in decline (Kostylev et al., 2003). Maps of the spatial variation in abiotic variables (e.g., salinity, temperature, sea color indices, bathymetry, sediment features) and epi-macrobenthos have been recognized as essential tools for sustainable management by facilitating monitoring of scale-dependent environmental fluctuations and anthropogenic pressures on benthic communities and habitats (Garrabou, Riera, and Zabala, 1998; Jackson, Attrill, and Jones, 2006; Siwabessy et al., 2000).
Benthic Classifications Using Bathymetric LIDAR
Mapping studies of benthic environments done over the past two decades have typically relied on physical sampling (i.e., grabs and dredges). This approach is not only time consuming and costly for the benthoscape (i.e., benthic habitats and their associated communities; Zajac, 2008) but also provides only scattered, discrete data across study areas. However, relatively recent advances in computer technologies, geographic information systems, underwater acoustic systems, and signal processing now provide effective tools to explore littoral patterns and processes as a complement to the physical sampling methods traditionally used to do benthic surveys (Durand, Legendre, and Juniper, 2006; Hutin, Simard, and Archambault, 2005). On the basis of relatively well known statistical patterns (Burns et al., 1989; Chivers, Emerson, and Burns, 1990; Clarke and Hamilton, 1999; Legendre et al., 2002), commercial bottom classifiers have been developed to allow researchers to extract seabed habitat information from returning acoustic signals (e.g., RoxAnn Systems). However, traditional shipboard survey methods have significant limitations, such as hazards, currents, and tides, which can be addressed by advances in airborne platforms (Guenther et al., 2000). In a single survey (,8 h), aircraft can cover a study area that would require a week or more for a vessel to survey. Aircraft provides seamless monitoring of water and terrestrial ecosystems without disturbances. The use of aircraft for bottom classification reduced to 10% (or less) the cost of that for shipboard surveys (Brown et al., 2002). Airborne LIDAR bathymetry (ALB) or airborne light detection and ranging (LIDAR) is a remote sensing technique for measuring water depths with an airborne scanning pulsed laser beam (Guenther, 1985). The technique is well suited to nearshore mapping because it provides the three-dimensional data needed to create an accurate digital terrain model (DTM) with 15-cm vertical accuracy (Irish, McClung, and Lillycrop, 2000). Compared with passive remote sensing systems, which are quite limited with respect to the depth to which they can measure, active LIDAR technology can measure depths up to three times Secchi depths, corresponding to about 60 m in very clear water (Guenther et al., 2000). The entire time history of the LIDAR return signal through the water column (i.e., waveform) is recorded so that the distance between the water surface and bottom can be measured. However, information contained in the waveform is often not fully exploited. Although the signal might be noisy because of optical sensors and combinations of environmental parameters, benthic waveform patterns related to seabed cover appear to be evident, suggesting that the laser signal might be used to characterize coastal benthic habitats. Innovations in ALB technology have allowed the nearshore environment to be mapped by draping intensity images over DTM, by combining intensity with passive data with the use of more sophisticated sensor or data fusion algorithms (Carter et al., 2001; Park et al., 2001; Tuell, 2002; Wang and Philpot, 2007) or by clustering benthic waveforms through statistical parameters (Collin, Archambault, and Long, 2008). High–spatial resolution (i.e., 1–10 m range) ALB imagery contains information that can be used to emphasize spatial patterns across benthoscapes (Zajac, 2008), such as areas of
87
breaks, similarity, and spatial autocorrelation. However, the classical approach of spectral analysis to identify discrete feature classes accounts only for correlation between spectral bands while neglecting correlation between neighboring pixels. Local spatial statistics, such as the local indicators of spatial autocorrelation (LISA) (Anselin, 1995) or Getis-Ord local Gi (Getis and Ord, 1992), look for specific areas in images that have clusters of similar or dissimilar values. Texture analysis, determined by variation in brightness, itself a function of uniformity, coarseness, regularity, frequency, and linearity (Musick and Grove, 1991), can also highlight local neighboring variability and thereby improve classification. Understanding and managing the benthoscape requires interdisciplinary studies and innovative methodologies. In this paper, the potential of using a LIDAR system to discriminate related sediment and benthic community features accurately is assessed for the dataset collected at Bonaventure, Gulf of St. Lawrence, Canada (Figure 1). The novelty of this approach lies in (1) comparing two classification procedures on the basis of multivariate analysis of LIDAR waveforms and (2) testing the significance of local spatial statistics and textural filtering outputs to delineate spatial patterns of benthic habitats.
SHOALS SYSTEM The Scanning Hydrographic Operational Airborne LIDAR Survey (SHOALS) system emits both 532- and 1064-nm wavelengths from a Nd:YAG laser (Guenther et al., 2000). The first wavelength (green) is intended for seabed detection because of its water penetration ability, whereas the second wavelength (infrared) allows sensing of the water surface and behaves as a near opaque surface at this wavelength. The transceiver records laser energy return time series (waveforms) with four detectors. One detector records the infrared energy reflected from the water surface (surface return), and two detectors collect the green energy reflected from the sea bottom (Geiger avalanche photodiode, GAPD; photomultiplier tube, PMT). A fourth detector records Raman energy at 645 nm (red) resulting from the excitation of water molecules at the water surface by the green laser energy (Pe’eri and Philpot, 2007; Wang and Philpot, 2007). Hence, Ramanchannel waveforms also can be used to calculate the air-water interface. The four receiving signals are digitized as four waveforms at a 1-ns sampling rate. Given the relative speed of light in air and water, the distance to the sea surface and bottom, and thus water depth, can be calculated for each bin (i.e., 1 ns) with a vertical resolution of 0.15 and 0.1125 m for air and water, respectively. The green laser produces about 7.5 mJ of light in a 6-ns pulse at a repetition rate of 3000 laser measurements per second. Because of eye safety regulations, the footprint size is approximately 2 m at the air-water interface. Because the highest laser measurement density was used in this study (i.e., a shot every 2 m), quasi–full coverage was achieved. SHOALS is a monostatic system such that the transmitter and the receiver are collocated and share the same field of view with a fixed nadir angle of 20u relative to the water
Journal of Coastal Research, Special Issue No. 62, 2011
88
Collin, Long, and Archambault
Figure 1. Location of the study area and map of the spatial distribution of ground-truthing in Bonaventure, Gulf of Saint Lawrence, in the south of Gaspesia Peninsula, Que´bec, Canada. Ground-truth stations sorted into four affinity groups (clustering of the dissimilarity matrix of epi-macrobenthos and sediments) are overlaid on a gray scale rasterization of elevation derived from SHOALS laser measurements (2-m resolution).
surface. The green laser pulse interacts with atmospheric and oceanographic factors, acting during the emitter-bottomdetector pathway (Figure 2a). Light propagation in water is much more altered than in air because of the enhanced extinction (ensemble of scattering and absorption processes). Interactions may be divided into three main groups: water surface, water column, and benthic (i.e., bottom) return (Guenther et al., 2000) (Figure 2b). The water surface interactions are determined by surface waves and specular reflection (Wang and Philpot, 2007). Within the water column, extinction from colored dissolved organic matter and both abiotic and biotic suspended particles constrain the laser beam to spread into an exponential cone (Figure 2a) (Guenther et al., 2000). Bottom interactions result from bottom rugosity and slope, as well as bottom reflectance at the pulse wavelength (a function of the composition of the benthos). A generic transmitted laser pulse is thus partially reflected from the water surface (2%) and from the sea bottom (4%–15%) to the airborne receiver (Guenther, 1985). All these interactions vary spatially and temporally. Because the altitude and all the loss parameters were constant during the survey, the expected backscatter intensity will follow Equation (1) (Guenther, 1985): PR ~WPT R|e({2KD) ,
ð1Þ
where PR is received power of the bathymetric LIDAR signal, W is constant combining loss factor, PT is transmitted power, R is benthic reflectance, K is the diffusive attenuation coefficient of water, and D is benthic depth. Each laser measurement from the SHOALS system is georeferenced with a real-time kinematic global positioning
system (RTK-GPS) to measure precisely the altitude of the aircraft with respect to the reference ellipsoid and aircraft position. The SHOALS system measures the distance from the aircraft to the sea bottom and the water surface and calculates the depth with respect to the reference ellipsoid. Studies by Irish, McClung, and Lillycrop (2000), Pope et al., (1997), and Riley (1995) have shown that the estimated accuracy of SHOALS by RTK-GPS is within 0.15 m in the vertical and 1 m in the horizontal planes, respectively. Roll, pitch, and yaw of the aircraft were also measured 200 times per second by an inertial measurement unit (IMU) to correct for changes in capture geometry.
METHOD Study Site The survey covered 82 km2 in 12 hours and consisting of 80 million individual depth and elevation measurements. They were collected by SHOALS between July 1 and 3, 2006, over the Baie des Chaleurs, southern Gulf of Saint-Lawrence, Que´bec, Canada. The maximum depth reached was 16.7 m over sandy zones and 11 m over macroalgae areas. LIDAR data for this work are focused on the subtidal nearshore of Saint-Sime´on-Bonaventure. This locality is hydrodynamically characterized by medium energy (Syvitski, 1992). Beaches of gravel and sand nourished by fluvioglacial deposits and an estuary dominated by the W-SW swell, are subject to twice daily tidal fluctuations that can reach 2 m. Whereas rocky areas (i.e., boulder and cobble) are mainly populated by the macroalgae Laminaria spp., sandy-gravely zones predominantly host the two polychaete Spionidae,
Journal of Coastal Research, Special Issue No. 62, 2011
Benthic Classifications Using Bathymetric LIDAR
89
Figure 2. (a) Air-water interface, water column, and bottom effects on the round pathway of the SHOALS green laser pulse (Guenther et al., 2000); (b) Segmentation of a generic bathymetric LIDAR backscatter. This one was acquired at 4.50 m depth. The oblique dashed line is a linear fit of the water column return.
Prionospio steenstrupi and Spiophanes bombyx; the bivalve Spisula spp.; the crustacean Corophium bonelli; and, finally, the echinoderm Echinarachnius parma, (‘‘Sand Dollar’’) (Long and Desrosiers, 2006). The study area was covered by a series of eight E–W overlapping flight lines, lasting less than 1 hour, at 269.3 6 9.6 m altitude, enabling a swath width of 198.1 6 4.4 m and a sample spacing of 2 m (i.e., ,2.5 million laser measurements covering 8 km2). Figure 3 is a shaded relief image of the DTM (i.e., a 3D rendering of grid of squares).
Underwater Ground-Truthing Within the extent of the survey, 339 stations specified the ground-truth information. Seafloor photographs were extracted with a digital high-resolution (5-megapixel) camcorder
Figure 3. DTM raster image of the study area (8 km2, 2-m resolution). The color ramp ranges from 2.05 to 10.91 m.
fitted with a wide-angle lens and placed in a waterproof case. Two 250 W light sources enabled adjustment of the illumination according to the water turbidity and the position of the camcorder to the bottom. The system was mounted on a cubic frame that included a reference ruler to evaluate the size of material on the seafloor. An underwater frame image of the seafloor covering an area of 0.16 m2 (0.4 3 0.4 m) was captured at all the stations in the study site. To determine the relevant utility of the laser data to discern differences in benthic characteristics, deviations in the composition of habitats, also called b-diversity, visualized with ground truth, was analyzed. These data were processed in two steps. First, to quantify the surface covered by the sediment and the epi-macrobenthos, a grid of 100 uniformly distributed points was superimposed on the photographs, and what was under each point was identified to give an estimate of the percentages of the surface covered by each component (Archambault, Banwell, and Underwood, 2001). Second, the aerial percentages were submitted to multivariate statistical analyses to classify the stations by their dissimilarity (PRIMER software; Clarke and Warwick, 1994). The matrix of 339 stations by 16 variables, corresponding to species and sediment type aerial percentages (Crustacea, Echinoidea, Annelida [and/or burrows], Gastropoda, Asteroidea, dead shells, Fucus spp., Zostera marina, Chondrus crispus, Laminaria spp., Chorda tomentosa, Polysiphonia spp., boul-
Journal of Coastal Research, Special Issue No. 62, 2011
90
Collin, Long, and Archambault
Table 1. The relative aerial density and contribution to within-group similarity for the macrobenthos and sediment aerial percentages in each group identified with hierarchical clustering.
Group
Features
Relative Aerial Density (%)
Contribution to WithinGroup Similarity (%)
G1
Laminaria spp. Pebbles .4 mm Fine sand .0.06 mm Pebbles .4 mm Cobbles .64 mm Laminaria spp. Zostera marina Fine sand .0.06 mm Pebbles .4 mm
73.2 11.7 87.1 40.6 26.3 11.2 45.3 29.9 16.2
97.98 6.98 98.7 51.67 32.41 7.64 53.98 29.5 13.09
G2 G3
G4
Figure 4. Underwater imagery (0.4 3 0.4 m) representing the four main habitat types classified by the affinity analysis. The selected habitats are (A) Laminaria spp. on pebble, (B) fine sand, (C) cobble-pebble with Laminaria spp., and (D) Z. marina on fine sand and pebble.
ders . 256 mm with encrusting algae, cobbles . 64 mm, pebbles . 4 mm, fine sand . 0.06 mm), was used to compute a dissimilarity matrix on selected variables with the BrayCurtis index (Clarke and Warwick, 1994). The dissimilarity matrix was then submitted to average linkage hierarchical clustering to classify the stations (Legendre and Legendre, 1998). Four affinity groups were identified from groupings of stations at the 65% level of dissimilarity (Figure 4). Assemblages at dissimilarities less than 65% were not examined because the number of stations per group was deemed to be too low. For each group, the mean relative density and percent contribution to within-group similarity (Table 1) were computed with the similarity percentage (SIMPER) routine to distinguish the components that had the greatest effect on the affinity groups identified (Clarke and Warwick, 1994). The four groups accordingly constituted habitats owning significant sediment and biological composition; namely, group 1 (G1) can be related to Laminaria spp. on pebble habitat, group 2 (G2) to fine sand habitat, group 3 (G3) to cobble-pebble with Laminaria spp. habitat, and group 4 (G4) to Z. marina on fine sand and pebble habitat. In summary, this cluster analysis gave robust results that are interesting to compare with the statistical analyses applied to the laser data. Within each habitat, the station closest to the mean was visually identified and considered as the most representative of the habitat. The photograph related to this station was then used to depict the ecological structure of this habitat. As a result, photographs A, B, C, and D in Figure 4 embody habitats G1, G2, G3, and G4, respectively. Seafloor sediments on the studied Bonaventure subset were either coarse with a dominance of cobbles-pebbles or fine sand. Cobbles (64–256 mm) and pebbles (4–64 mm) were the most common sediment (up to 78%) in the open sea from the NW to the SE of the area. Fine sand coverage (0.06–4 mm),
located on the closer part of the beach along coastline, met 86%. Distinct benthic communities were identified from the analysis of the photographs, and they tended to be distributed over peculiar sediment features. The most frequent taxa belong to the Plantae kingdom, namely Laminaria spp. and Z. marina. Laminaria spp. were very abundant, especially on cobbles, rarely overturned, on which its claw-like holdfasts settle. In contrast, Z. marina commonly covered fine sand and rarely pebbles.
Signal Processing of Return Waveform The integration of environmental parameters acting at the air-water interface and underwater, as well as the variability of electronic instruments within the green waveform, prompted a denoising method. The fast Fourier transform (FFT) algorithm decomposed the one-dimensional time series into components of different frequencies. The FFT routine and a low-pass filter (Research Systems, 2005) were applied to the waveform for further analysis as a denoised signal. The shape and power of the returned signal from the seafloor can change significantly with seafloor depth as a result of scattering and absorption, even if the seafloor habitat remains the same (Clarke and Hamilton, 1999; Hamilton, 2001). To overcome this, the statistical variables, describing the bottom waveform, were regressed on depth, and the residuals constituted the only variance analyzed. Because the green waveform is a logarithmic digitization of photo counts, as revealed by the linear behavior of the water column, the bottom signal accordingly appears nonlinear. Therefore, the regression consisted of a nonlinear least squares fit, called the Gaussfit function (Research Systems, 2005), based on a linear combination of Gaussian and quadratic functions with six different terms: 2 f ð xÞ~A0e{½ðx{A1Þ=A2 =2 zA3zA4xzA5x2 ,
ð2Þ
in which A0, A1, A2, A3, A4, and A5 are the unknown parameters. LIDAR data were recorded in two file formats according to their information. The LAS file format contains the attributes easting, northing, ellipsoidal height, intensity, GPS time, flight line number, and date. The INW file format conveys the four waveforms and the timestamp. Both file formats were
Journal of Coastal Research, Special Issue No. 62, 2011
Benthic Classifications Using Bathymetric LIDAR
Table 2.
91
Parametrization of nonlinear regressions applied to 12 extracted variables and Pearson’s correlation coefficient between water depth and residuals.
Variables
A0
A1
A2
A3
A4
A5
Mean Variance Skewness Kurtosis Median Mean absolute deviation Area Ampl. S/B* Mean curve Variance curve Skewness curve Kurtosis curve
20.4471 24.7086 20.054 20.0574 20.5019 211.7151 2714.532 0.5007 270.1352 69.1975 0.4838 20.238
25.209 24.0577 23.19 26.4351 24.454 22.3444 23.5614 21.3418 0.0979 25.4541 22.6 22.5661
21.4477 20.1839 0.0212 2.1814 20.5453 0.5527 23.7595 20.1372 0.1654 1.9827 0.6711 0.2051
153.2004 384.8938 20.0299 21.8083 153.8885 24.2512 2734.7187 0.5441 162.3059 120.1035 21.1463 20.868
10.7133 76.2365 0.2367 20.4235 9.5514 4.497 2393.457 21.3592 12.1691 63.5209 20.561 0.0168
0.0184 6.3106 0.0156 20.0296 20.0694 0.331 258.5402 0.7274 0.1823 5.6655 20.0456 20.0013
Correlation Depth vs. Residuals
1 25 1 3 5 24 3 22 2 1 9 8
3 3 3 3 3 3 3 3 3 3 3 3
1022 1023 1026 1022 1023 1024 1023 1023 1024 1023 1023 1024
* Ampl. S/B 5 Ratio Amplitude Surface / Amplitude Bottom.
bridged for each LIDAR point with the keystone GPS time timestamp (Cottin, 2008). The data were processed into IDLENVI 4.4 (Research Systems, 2005) point coverages, and different surfaces were constructed. The ellipsoidal heights (provided by an Optech custom algorithm), as well as the 12 descriptive statistics, were used to construct triangular irregular networks (TINs) that were linearly interpolated to a 2-m DTM and 12, 2-m grid images, respectively. Then, the signal processing consisted of extracting the portion of the waveform that encompassed relevant benthic information. The portion of interest, called ‘‘benthic waveform’’ (Figure 2b), was retrieved in examining the signal curvature through the first derivative. More precisely, the first step consisted of pruning the waveform between (1) the bin of the surface return (first derivative) + 20 ns, so that the influence of surface was deleted, and (2) 150 ns, located after the informational portion and before the artifact value at the end of the derivative. The impending step aimed to delineate the signal of interest by the previous lower bound and the new minimum, corresponding to the inflection point of the slope after the bottom. The final step focused on the new maximum, corresponding to the inflection point of the bottom return. After that, a series of descriptive statistical variables were computed—namely, mean, variance, skewness, kurtosis, median, mean absolute deviation, area under curve, amplitude difference between bottom and surface returns (derived from the benthic waveform), and mean, variance, skewness, and kurtosis describing the curve placed between the end of the water column and the beginning of the benthic portion. These parameters were chosen owing to their common use for acoustic bottom characterization (Durand, Legendre, and Juniper, 2006). This yielded a table of 12 extracted variables (columns) and ,2.5 million laser measurements (rows). Every variable of this matrix was first nonlinearly regressed on depth according to parameters identified in Table 2. The efficiency of this regression was supported by the absence of correlation (Pearson’s coefficient) between residuals and depth (last column of Table 2). Then, residuals were converted into 12 PCs with a principal component analysis (PCA) (Legendre and Legendre, 1998). For sorting out the benthic backscatter in SHOALS, the first PC band and the succeeding PC bands, with a sum of eigenvalues was greater than 90% of
the total variance in the dataset, were chosen (Legendre et al., 2002). Within the remote sensing context, the PCA is intended for producing uncorrelated output bands, segregating noise components, and reducing the dimensionality of the dataset. The scores on the 12 PCA components were submitted to one classical unsupervised and one state-ofthe-art supervised classifiers, namely K-means and support vector machine (SVM). These two methods were chosen to evaluate the relevance of the training algorithm by quantifying the performance gained between the unsupervised and supervised processes. Both algorithms were implemented within IDL-ENVI 4.4 (Research Systems, 2005).
Benthic Waveform Classifications The K-means cluster analysis performed an iterative alternating fitting process, and the optimal split level was determined by the number of classes resulting from the ground-truthing. In essence, the SVM classification is based on the notion of fitting an optimal separating hyperplane between classes by focusing on the training samples that lie at the edge of the class distributions, the support vectors. All of the other training samples are effectively discarded because they do not contribute to the estimation of hyperplane location (Foody and Mathur, 2004). Thus, a limited number of spectrally representative training samples (e.g., three of four per class) may be used to classify a dataset as accurately as a larger (two- to threefold) training set derived in a conventional manner. The set of pixels from the DTM and 12 statistical grid images, representing the discriminated ground-truth stations, was divided into a training set, for classifier calibration, and a validation set, to provide an independent test of classification performance. This information was summarized by the overall accuracy, A, defined as the ratio of the number of validation pixels that are classified correctly to the total number of validation pixels irrespective of the class (Belluco et al., 2006). A further important confusion matrix statistic used here is the Kappa coefficient, K, which describes the proportion of correctly classified validation sites after random agreements are removed (Rosenfield and Fitzpatrick-Lins, 1986).
Journal of Coastal Research, Special Issue No. 62, 2011
92
Collin, Long, and Archambault
Geostatistical Analysis To overcome the limitation of the global spatial autocorrelation statistics, local indicators of spatial association (LISAs) have been developed to focus on local variations within patterns of spatial dependence, resulting in a potential to uncover discrete spatial regimes that might be overlooked by existing ‘‘global’’ techniques (Anselin, 1995). Three local spatial statistics were chosen for their significance in remotely sensed imagery—namely, Local Moran’s I (LI), Getis-Ord Local Gi (LGi), and Local Geary’s C (LC) (Anselin, 1995; Getis-Ord, 1992). Whereas LI identifies pixel clustering, LGi underlines hot spots, such as areas of very high or very low values that occur near one another; finally, LC allows characterization of areas of high variability between a pixel value and its neighboring pixels. Basically, these LISAs appear as proxies for quantifying the heterogeneity between neighboring pixels by finding clusterings (similar values plus their sign) and detecting edges (high variability). Texture refers to the spatial variation of image tone as a function of scale. To be defined as a distinct textural area, the gray levels (GLs) within the area must be more homogeneous as a unit than areas having a different texture. The following analysis evaluated the contribution of several textural filters, on the basis of occurrence and cooccurrence measures to the improvement of discrimination in the 12 digital models, derived from SHOALS waveforms parameters. Occurrence represents the frequency of occurrence of each GL within the processing window. Four different occurrence filters were coded: mean (Mo), variance (Vo), skewness (So) and entropy (Eo) (Anys et al., 1994). The first three filters are descriptive filters, whereas the last filter represents orderliness. Co-occurrence measures correspond to the frequency of occurrence of each pair of GLs in the window. The GL cooccurrence matrix (GLCM), also called the gray tone spatial dependency matrix, is a matrix of relative frequencies with which pixel values occur in two neighboring processing windows separated by a specified distance and direction. It shows the number of occurrences of the relationship between a pixel and its specified neighbor. Derived filters of the GLCM included mean (Mco), variance (Vco), correlation (CORco), contrast (CONco), dissimilarity (Dco), entropy (Eco), and angular second moment (Aco) (or energy) (Anys et al., 1994; Haralick et al., 1973). These implemented filters compose groups of descriptive (first three), contrast (fourth and fifth), and orderliness (last two) features. Success of the classification procedure with textural features depends on the choice of ad hoc distance between pixels and the window size. Because a natural scene is composed of a multitude of textures with quite variable degrees of fineness or coarseness, a distance equal to one pixel was adequate for both fine and coarse textures. The selection of the window size had to be optimal to enhance and retrieve spatial patterns of interests. By means of the coefficient of variation (i.e., the standard deviation divided by the mean) of a given feature for each benthic class with respect to window size (Anys and He, 1995), the adequate window size elected corresponded to 5 3 5 pixels of the 12 digital models.
Figure 5. Plot of the eigenvalue as a function of the cumulated percentage of eigenvalues issuing from the forward principal component (PC) rotation applied to the 12 parameter bands. The dashed line shows evidence for the relevance for taking the first four PCs.
Contribution of the Local Spatial Statistics and the Textural Features An approach for evaluating properties consists of calculating a feature on the image and establishing some distance functions between classes in such a way that the larger the distance, the more accurate will be the classification. The Jeffries-Matusita (JM) distance is one of the most popular statistical separability measurements in remotely sensed data analysis. The JM distance between a pair of probability functions is the measure of the average distance between the two class density functions (Richards, 1993; Swain and Davis, 1978). For normally distributed classes, this distance becomes the Bhattacharyya distance (Richards, 1993). For an exclusive discrimination, the JM distance reaches 2, whereas it is 0 for a complete similarity. Hence, for each of the 12 digital models, the three local spatial statistics and the 11 textural features were applied; then, samples of every benthic class in output images were surrogates used to calculate the JM distances. By quantifying the difference between distances for each benthic type in feature space before and after the addition of the geostatistical features, their contribution could be calculated.
RESULTS PCA Applied to SHOALS Benthic Parameters Because multispectral data bands are often highly correlated, the PC transformation is intended for building up uncorrelated output bands. The PC method provides a new set of orthogonal axes that have their origin at the data mean and that are rotated so that the data variance is maximized (Legendre and Legendre, 1998). The first four PC bands (Figure 5), representing 90.41% of the whole initial variance, acutely reduced the number of bands to process (divided by 3) while preserving the multiparameter information. The first PC bands (up to four or five) contained the largest percentage of data variance,
Journal of Coastal Research, Special Issue No. 62, 2011
Benthic Classifications Using Bathymetric LIDAR
93
Figure 6. Resulting images of the digital terrain model (a), and the 12 statistical parameters derived from SHOALS benthic waveforms classified by the Kmeans (b) and the support vector machine (SVM) (c) algorithms. Resulting images of the SVM classifier applied to the 12 statistical parameters plus the mean derived from the GLOM, the mean derived from the GLCM, and the Getis-Ord Local Gi (d) and applied to the first four PCs plus the mean derived from the GLOM, the mean derived from the GLCM, and the Getis-Ord Local Gi (e).
whereas the last PC bands had very little variance, much of which is due to noise in the original spectral data.
Benthic Waveform Classifications On the basis of statistical discrimination of the four habitats identified in the ground-truth work, their inherent region of interests (ROIs) allowed us to discard or accept classifications. As a result, the number of clusters for the unsupervised classification was logically determined for K 5 4 (one additional class is used as a null class attributed to the surrounding study area), and the process converges at the 18th iteration. The results of the benthic classifications of both algorithms are visualized in Figure 6b and 6c and given in Table 3, which highlights the characteristics summarizing the ‘‘validation’’ confusion matrices for K-means and SVM.
Table 3. Overall Accuracy (A, %) and Kappa coefficient (K) summarizing the confusion matrix characteristics for both classifications and validation sites considered. G1 A
K-means SVM
G2 K
A
G3 K
A
G4 K
A
Total K
A
K
31.1 0.24 79.3 0.69 77.8 0.73 82.8 0.79 67.7 0.61 78.3 0.83 94.6 0.91 97.8 0.89 88.9 0.82 89.9 0.86
The implemented extraction algorithm applied to the 12 bands derived from the benthic waveform revealed an overall robust classification performance, especially for SVM, as the assessment clues stressed it (K-means: A 5 67.7%, K 5 0.61; SVM: A 5 89.9%, K 5 0.86). On the one hand, the fine sand (G2) habitats, the cobblepebble with Laminaria spp. (G3), and Z. marina on fine sand and pebble (G3) were identified correctly (A 5 79.3%, K 5 0.69; A 5 77.8%, K 5 0.73; A 5 82.8%, K 5 0.79, respectively) by K-means clustering. The last habitat, Laminaria spp. on pebble (G1), was characterized by poorer performances of the unsupervised classification both statistically (A 5 31.1%, K 5 0.24) and visually. Running the supervised classifier over the same statistical parameters provided a better performance of classification. Validation sites related to G2 and G3 were very well classed (A 5 94.6%, K 5 0.91; A 5 97.8%, K 5 0.89, respectively), whereas the statistics describing G4 and G1 showed lower performance, albeit relatively high (A 5 88.9%, K 5 0.82; A 5 78.3%, K 5 0.83, respectively). Even though the statistics resulting from the confusion matrix indicated a better performance of the algorithm SVM than the K-means, both classifications were in satisfactory agreement about the structural nature of benthic habitats. Indeed, G2 and G3 were constituted by substrata characterized by a low spectral variability (i.e., fine sand and cobble-pebble),
Journal of Coastal Research, Special Issue No. 62, 2011
94
Collin, Long, and Archambault
of separability among habitats were averaged in relation to both band combinations (Table 5). Through a global perspective (total with 12 bands), the best local spatial and textural features selected were the LGi for the spatial statistics (0.98), Mo for the GLOM (1.40), and Mco for the GLCM (1.49). Other local spatial statistics had approximately the same poor performance of discrimination (0.14 and 0.10). Concerning the GLOM, the mean (1.40) was followed by the variance (0.55), entropy (0.14), and skewness (0.13). For the GLCM, the mean (1.49) led contrast (0.74) and variance (0.52) and highly outperformed the remaining features: dissimilarity (0.25), correlation (0.19), entropy (0.12), and energy (0.10). Within the framework of the three best features, a hierarchy was also outlined at the habitat scale. G2 (1.07, 1.81, 1.82) was well differentiated, closely followed by G3 (1.02, 1.66, 1.61), G4 (0.91, 1.21, 1.37), and finally G1 (0.93, 0.93, 1.18). The same patterns of separability arose from the first four PCs. On the one hand, for the totals, LGi (0.95), Mo (1.36), and Mco (1.46) were the most efficient features in terms of benthic discrimination, whereas among the four habitats, the sorting of separability values was consistent with that of the 12 bands combination: G2 (1.04, 1.77, 1.78), G3 (1.00, 1.61, 1.58), G4 (0.88, 1.15, 1.32), and G1 (0.90, 0.92, 1.15).
Table 4. Overall mean of the JM distances between the four benthic habitats for both band combinations: 12 bands (all bands), first four principal components (90.41% of the initial variability). Every number was rounded up to the second decimal.
Twelve bands First four PCs
G1
G2
G3
G4
Total
1.55 1.28
1.82 1.67
1.73 1.59
1.61 1.41
1.68 1.49
whereas the greater heterogeneity of G1, because of the threedimensional structure of macroalgae, might entail misclassifications with the other habitat classes. Despite the vegetation cover, G4 was well discriminated by both algorithms, witnessing the spectral homogeneity of these intertidal meadows.
Contribution of the Local Spatial Statistics and Textural Features Aiming to evaluate the contribution of the local spatial statistics and textural filters, class separability was computed from both the 12 bands and the first four PCs. Separability was calculated by choosing training ROI for each benthic habitat and by computing the mean of JM distance for each benthic habitat with respect to both band combinations (Table 4). In addition, the overall mean per combination was estimated (‘‘Total’’ column) to provide an indication of the performance of each combination in terms of benthic separability. Results of 12 bands (1.68) showed that this fully informational combination offers a slightly better overall discrimination than the four PCs (1.49). These overall performances tended to corroborate the downward trend of the information (percentage of the initial variability) within the four first PCs (Figure 5). The hierarchy of the overall performance (12 bands . four first PCs) was also respected at the habitat scale. Furthermore, G2 and G3 were especially well discriminated with respect to G1 and G4 for the 12 bands and the first four PC combinations. Then, local spatial and textural features were applied to each of the 12 bands and each of the first four PCs (hence, 224 [5 16 bands 3 14 features] images generated) and mean rates
Integration of the Best Local Spatial Statistics and Textural Features for Benthic Classification A merging approach, integrating the best features applied to each ‘‘raw’’ parameter, might improve the power of benthic habitat discrimination. Mean rates of separability issued from the coalescence of the features were calculated for both band configurations (Table 6). For all habitats (i.e., the Total column), both ‘‘boosted’’ combinations revealed separability rates greater than previous combinations (Table 4). Hence, the overall performance gained 0.10 with 12 bands and 0.21 with the first four PCs. As a result, the overall accuracy (A) quantifying the performance of the supervised classification increased to 95.2% and 93.3% for 12 bands and four PCs,
Table 5. Overall mean of the JM distances between the four benthic habitats calculated for each spatial and textural feature for both band combinations: 12 bands and first four principal components (90.41% of the initial variability). Every number was rounded up to the second decimal. G1
LI LGi LC Mo Vo So Eo Mco Vco CORco CONco Dco Eco Aco
G2
G3
G4
Total
12 Bands
4 PCs
12 Bands
4 PCs
12 Bands
4 PCs
12 Bands
4 PCs
12 Bands
4 PCs
0.03 0.93 0.05 0.93 0.28 0.02 0.08 1.18 0.23 0.03 0.74 0.09 0.05 0.04
0.02 0.90 0.05 0.92 0.28 0.02 0.07 1.15 0.21 0.03 0.69 0.08 0.05 0.04
0.28 1.07 0.17 1.81 0.83 0.33 0.26 1.82 0.85 0.37 0.79 0.43 0.22 0.15
0.26 1.04 0.16 1.77 0.81 0.32 0.25 1.78 0.83 0.35 0.75 0.42 0.22 0.14
0.21 1.02 0.09 1.66 0.78 0.12 0.13 1.61 0.73 0.29 0.75 0.32 0.15 0.17
0.20 1.00 0.09 1.61 0.76 0.11 0.13 1.58 0.71 0.28 0.72 0.30 0.15 0.16
0.05 0.91 0.11 1.21 0.32 0.07 0.11 1.37 0.28 0.09 0.68 0.16 0.07 0.06
0.05 0.88 0.11 1.15 0.29 0.07 0.10 1.32 0.27 0.09 0.65 0.15 0.07 0.05
0.14 0.98 0.10 1.40 0.55 0.13 0.14 1.49 0.52 0.19 0.74 0.25 0.12 0.10
0.13 0.95 0.10 1.36 0.53 0.13 0.14 1.46 0.50 0.19 0.70 0.24 0.12 0.10
Journal of Coastal Research, Special Issue No. 62, 2011
Benthic Classifications Using Bathymetric LIDAR
Table 6. Overall mean of the JM distances between the four benthic habitats for both band combinations plus the mean derived from the GLOM, the mean derived from the GLCM, and the Getis-Ord Local Gi. Every number was rounded up to the second decimal.
Twelve bands + three features Four PCs + three features
G1
G2
G3
G4
Total
1.71
1.91
1.75
1.77
1.78
1.52
1.87
1.78
1.66
1.70
respectively (Figure 6d and 6e). The superiority of four PCs compared with 12 bands was confirmed for all habitats: G1 (0.24 vs. 0.16), G2 (0.20 vs. 0.09), G3 (0.19 vs. 0.02), and G4 (0.17 vs. 0.16). However, the enhanced rates of distinction were appreciably the same among the four habitats, except for G2 and G3 in the 12-bands combination, which appeared lower. The specific contribution of the statistics and filters could readily be estimated in deducting the initial mean rates of separability from the enhanced combinations (Table 7). For the 12 bands, Mco and Mo both exhibited the highest overall contribution values. Nevertheless, these previous features did not provide the best contributions for most benthic habitats. Indeed, both contributions only predominated for G2 and G3 (Mo: 0.39 and 0.25; Mco: 0.31 and 0.27, respectively). The local spatial index, LGi, enabled the enhancement of the other two habitats, G1 and G4 (0.32 and 0.21, respectively). Following the pattern of Table 6, contributions of the four PCs gave slightly better results than those of the 12 bands. Both overall contributions were attributed to Mco (0.30) and Mo (0.28). Additionally, regarding the pattern resulting from 12 bands, Mco and Mo constituted the most substantial contributions for G2 and G3, whereas LGi reached the highest contribution for the two benthic habitats G1 and G4.
DISCUSSION Abiotic components influence the ecosystem function (Turner, Gardner, and O’Neill, 2001). As such, bottom rugosity (i.e., ratio of the benthoscape surface area to the planimetric area) is linked with the distribution of benthic species (Brock et al., 2006). Because SHOALS can quantify rugosity at the meter scale, it would be logical to test this benthoscape index as a surrogate for discriminating benthic communities and thus classifying their habitats. Furthermore, the variability of the seafloor bathymetry influences benthic community structure and ecological processes at many spatial scales (Collin, Long, and Archambault, 2008b; Cusson and Bourget, 1997; Guichard and Bourget, 1998) and modifies diversity of species within the benthic community (Archambault and Bourget, 1996). Therefore, a 1-m DTM enhanced by adequate local spatial and textural analyses can be used to detect biodiversity anomalies. The enhanced DTM also can be used with other factors to characterize habitats, such as sediment types, seafloor geomorphology, salinity, turbidity, temperature range, current speed, food supply, predation and competition pressure, disturbance by fishing, and other anthropogenic activities (Kostylev et al., 2003).
95
Table 7. Overall mean contributions of the three best discriminant features between the four benthic habitats for both band combinations. Every number was rounded up to the second decimal. G1
G2
G3
G4
Total
0.32 0.15 0.19
0.09 0.39 0.31
0.14 0.25 0.27
0.21 0.12 0.14
0.19 0.23 0.23
0.41 0.22 0.27
0.14 0.43 0.38
0.16 0.27 0.34
0.23 0.19 0.22
0.23 0.28 0.30
Twelve bands LGi Mo Mco Four PCs LGi Mo Mco
These environmental factors produce substantial spatial variations among benthic habitats and have to be considered when interpreting SHOALS ALB classifications. Spatial variations of the bathymetry have been shown to impede differentiation between remotely sensed signatures of the benthos (Durand, Legendre, and Juniper, 2006; Hutin, Simard, and Archambault, 2005; Legendre et al., 2002). One of the major issues within the analysis of the ALB waveform is the use of depth normalization procedures (Collin, Archambault, and Long, 2008). A depth normalization procedure recognizes the benthic assemblages at different scales, ranging from centimeters to kilometers in deep waters (Jumars and Ekman, 1983) and in the intertidal zone (Archambault and Bourget, 1996; Underwood and Chapman, 1998). The depth regression removed the depth-related variation from each benthic parameter and generated satisfactory classifications independent of the classification type (unsupervised or supervised). The depth regression was applied to the statistical parameters derived from the depthskewed bottom waveform. However, the spatial and temporal contribution to the bottom return can be retrieved only by employing an analytical procedure before extraction of statistical parameters. The value of the signal returned at a given seafloor depth (D) was corrected for the water attenuation and normalized to a reference seafloor depth (D0). The benthic parameters were extracted at this reference seafloor depth. In addition to the depth effect, it was noticeable that other factors affected the performance of the discrimination of bottom type with the use of a depth-corrected bottom return amplitude. These factors included the viewing angle of the SHOALS ALB system relative to bottom and water surface conditions (Wang and Philpot, 2007). Further laboratory experiments will investigate how bottom type affect the time and amplitude of bottom return signals. These laboratory results will enable us to outline the values of the bio-optical parameters numerically, to be implemented in the radiative transfer model for ALB. Classification results in this study showed that the supervised classifier, SVM, outperformed the unsupervised algorithm (K-means). These results corroborate the consensus that training a classifier can gather the entire variability of GLs defining a class. Moreover, the SVM algorithm requires only the training samples to support the vectors. It is important to note that the training samples are located at
Journal of Coastal Research, Special Issue No. 62, 2011
96
Collin, Long, and Archambault
the edge of the class distribution in feature space because all other training samples do not provide a more robust classification analysis (Foody and Mathur, 2004). Consequently, the support vectors are likely to be identified from the ROI, and accurate classification of shallow seabed habitats can be conducted solely with the use of a small training dataset. Because ground-truthing is time consuming (e.g., visual inspection and physical sampling), the use of the SVM classifier might be advocated to reduce the duration of this step while keeping high rates of classification. An integration of spatial patterns, as underlined by the previous spatial-dependent results, will allow the operator to distinguish better between the benthic classes. Therefore, an object-oriented segmentation that takes into account the spatial context might outperform traditional classification based on statistics at the scale of single pixels and might also yield a reliable restitution of the shallow-water seabed. The overall mean values of the JM distances between the four benthic habitats for both band combinations without spatial or textural features (Table 4) indicated a significantly better discrimination of the two bare habitats (G2 and G3) compared with the two vegetated habitats (G4 and G1). This discrimination showed no correlation to grain-size or bottom morphology. The rates of separability tended to decrease with an increase in vertical heterogeneity over the seabed. The most discriminated habitat was the relatively flat fine sand (G2), closely followed by cobbles-pebbles with Laminaria spp. (G3) with a sizeable degree of rugosity; Z. marina with fine sand and pebbles (G4) was the third habitat distinguished, displaying a more complex three-dimensional structure, and finally, the habitat dominated by 3-m-high macroalgae, Laminaria spp., on pebbles (G1), exhibiting the most complex structure. Before computing the features presented in Table 4, the hierarchy of separability among habitats was calculated with respect to the three best features (Table 5). The two best textural features for calculating the hierarchy of separability were the mean values on the basis of the frequency of occurrence of either each GL or each pair of GLs, Mco and Mo, derived from the GLOM and GLCM, respectively. The results from Mco and Mo allowed better differentiation of bare habitats (G2 and G3) from habitats covered with vegetation (G4 and G1). Basically, smoothing the 12 ‘‘raw’’ parameters can keep the global performance of benthic classification while mitigating the computation time. However, the local spatial statistics showed that the best homogeneous rate of separability for the four habitats was the LGi, which revealed ‘‘GL hotspots’’ (i.e., the concentration of very high or very low GL occurring close to one another). From Table 5, it was also noticed that the LGi better discriminated vegetated habitats compared with the ability to discriminate bare habitats (Table 7). Furthermore, the CONco textural feature (contrast issued from the GLCM) generated JM distances (in Table 5) rigorously equivalent to those derived from LGi. This feature relies on the difference between the highest and lowest GL of a contiguous set of pixels, which is why CONco acted as LGi and reached the fourth best feature. In Tables 5 through 7, systematic comparisons between the 12 bands (entire information) and the first four PC combina-
tions (90.41% of initial variance) suggested that the four PC results corroborated the results received from the 12 bands. Although use of the first four PCs did not represent all the information retained (.90%), the performance of the classification results was markedly conserved while the number of images to process fell from 168 to 56, considerably diminishing the processing time.
CONCLUSIONS The study results included benthic habitat classifications, local spatial statistics, and textural features. The main conclusions were as follows: (1) the multivariate statistical processing of ALB SHOALS benthic waveforms enabled a correct recognition of benthoscape patterns with acceptable rates of classification. The classification results were supported by the underwater ground-truth clustering with an 89.9% and 67.7% rate of classification, obtained for SVM (supervised classification) and K-means (unsupervised classification), respectively. (2) The best discrimination values were provided from the assessment of local spatial statistics and textural filters, illustrated by the Getis-Ord Local Gi (mean derived from the occurrence matrix and mean derived from the co-occurrence matrix). (3) The local spatial statistics and textural filters with the 12 benthic parameter bands improved the separability of benthic habitats, and the rate of correct (supervised) classification increased by 5.3%. (4) The local spatial statistics and textural filters also provide insights into biophysical aspects of the habitat. The two textural filters, Mco and Mo, served as surrogate estimators of sediment habitats in spectrally homogeneous sites. The local spatial statistics, LGi and the contrast issuing from the co-occurrence matrix, CONco, provided robust indicators of vegetated habitats, characterized by a complex three-dimensional structure. (5) The use of the first four PCs, presenting 90.41% of the initial variance, led to the same results as the 12 bands classification. The rate of classification of the four synthesized PCs with the three previous features equal to 93.3%, while dramatically gaining computation time.
ACKNOWLEDGMENTS A. Collin thanks David Streutker, Kevin Lausten, and Antoine Cottin for their IDL-ENVI modules implemented for LIDAR advancements. Chris McKindsey and Nancy Lamontagne, anonymous reviewers, and the editors of this special session are acknowledged for their relevant corrections and suggestions.
LITERATURE CITED Anselin, L., 1995. Local indicators of spatial association—LISA. Geographical Analysis, 27(2), 93–115. Anys, H. and He, D.C., 1995. Evaluation of textural and multipolarization radar features for crop classification. IEEE Transactions on Geoscience and Remote Sensing, 33(5), 1170–1181. Anys, H.; Bannari, A.; He, D.C., and Morin, D., 1994. Texture analysis for the mapping of urban areas using airborne MEIS-II images. International Journal of Remote Sensing, 19(5), 883–894. Archambault, P. and Bourget, E., 1996. Scales of coastal heterogeneity and benthic intertidal species richness, diversity and abundance. Marine Ecology Progress Series, 136, 111–121.
Journal of Coastal Research, Special Issue No. 62, 2011
Benthic Classifications Using Bathymetric LIDAR
Archambault, P.; Banwell, K., and Underwood, A.J., 2001. Temporal variation in the structure of intertidal assemblages following the removal of sewage. Marine Ecology Progress Series, 222, 51–62. Belluco, E.; Camuffo, M.; Ferrari, S.; Modenese, L.; Silvestri, S.; Marani, A., and Marani, M., 2006. Mapping salt-marsh vegetation by multispectral and hyperspectral remote sensing. Remote Sensing of Environment, 105, 54–67. Brock, J.C.; Wright, C.W.; Kuffner, I.B.; Hernandez, R., and Thompson, P., 2006. Airborne lidar sensing of massive stony coral colonies on patch reefs in the northern Florida reef tract. Remote Sensing of Environment, 104, 31–42. Brown, E.D.; Churnside, J.H.; Collins, R.L.; Veenstra, T.; Wilson, J.J., and Abnett, K., 2002. Remote sensing of capelin and other biological features in the North Pacific using lidar and video technology. ICES Journal of Marine Science, 59, 1120–1130. Burns, D.R.; Queen, C.B.; Sisk, H.; Mullarkey, W., and Chivers, R.C., 1989. Rapid and convenient acoustic seabed discrimination for fisheries applications. Proceedings of the Institute of Acoustics, 11, 169–178. Carter, W.; Shrestha, R.; Tuell, G.; Bloomquist, D., and Sartori, M., 2001. Airborne laser swath mapping shines new light on earth’s topography. Eos, Transactions, American Geophisical Union, 82(46), 549–555. Chivers, R.C.; Emerson, N., and Burns, D.R., 1990. New acoustic processing for underway surveying. Hydrographic Journal, 56, 9– 17. Clarke, K.R. and Warwick, R.M., 1994. Similarity-based testing for community pattern—the 2-way layout with no replication. Marine Biology, 118(1), 167. Clarke, P.A. and Hamilton, L.J., 1999. The ABCS Program for the Analysis of Echo Sounder Returns for Acoustic Bottom Classification. DSTO-GD-0215. Melbourne, Australia: Defence Science and Technology Organisation, Aeronautical and Maritime Research Laboratory. Collin, A.; Archambault, P., and Long, B., 2008a. Mapping the shallow water seabed habitat with the SHOALS. IEEE Transactions on Geoscience and Remote Sensing, 46(10), 2947–2955. Collin, A.; Long, B., and Archambault, P., 2008b. Relating spatial benthoscape patterns with a high resolution bathymetric LiDAR. In: Proceedings of the 31st International Conference of Coastal Engineering—ICCE (Hamburg, Germany), 31 August–5 September, pp. 4645–4657. Costanza, R.; d’Arge, R.; de Groot, R.; Farber, S.; Grasso, M.; Hannon, B.; Limburg, K.; Naeem, S.; O’Neill, R.V.; Paruelo, J.; Raskin, R.G.; Sutton, P., and van der Belt, M., 1997. The value of the world’s ecosystem services and natural capital. Nature, 387, 253–260. Cottin, A., 2008. Adaptation du SHOALS-3000T a` la cartographie se´dimentaire des zones peu profondes de la coˆte nord de la Baie des Chaleurs, golfe du Saint-Laurent, Que´bec, Canada. Que´bec, Canada: L’Institut national de la recherche scientifique Eau Terre Environnement. Doctoral thesis. Cusson, M. and Bourget, E., 1997. Influence of topographic heterogeneity and spatial scales on the structure of the neighboring intertidal endobenthic macrofaunal community. Marine Ecology Progress Series, 150, 181–193. Durand, S.; Legendre, P., and Juniper, S.K., 2006. Sonar backscatter differentiation of dominant macrohabitat types in a hydrothermal vent field. Ecological Applications, 16(4), 1421–1435. Evans, D.L.; Alpers, W.; Cazenave, A.; Elachi, C.; Farr, T.; Glackin, D.; Holt, B.; Jones, L.; Liu, T.W.; McCandless, W.; Menard, Y.; Moore, R., and Njoku, E., 2005. Seasat-A 25-year legacy of success. Remote Sensing of Environment, 94(3), 384–404. Foody, G.M. and Mathur, A., 2004. Toward intelligent training of supervised image classifications: directing training data acquisition for SVM classification. Remote Sensing of Environment, 93, 107– 117. Gaffin, S.; Engelman, R., and Hachadoorian, L., 2006. Mapping the future of world population. Poster. New York: Center for Climate Systems Research, Columbia University. Garrabou, J.; Riera, J., and Zabala, M., 1998. Landscape pattern indices applied to Mediterranean subtidal rocky benthic communities. Landscape Ecology, 13, 225–247.
97
Getis, A. and Ord, J.K, 1992. The analysis of spatial association by use of distance statistics. Geographical Analysis, 24(3), 189–206. Guenther, G.C., 1985. Airborne Laser Hydrography: System Design and Performance Factors. NOAA Professional Paper Series, March. Rockville, MD: NOS, NOAA. Guenther, G.C.; Cunningham, A.G.; LaRocque, P.E., and Reid, D.J., 2000. Meeting the Accuracy Challenge in Airborne LiDAR Bathymetry. Dresden: European Association of Remote Sensing Laboratories. Guichard, F. and Bourget, E., 1998. Topographic heterogeneity, hydrodynamics, and benthic community structure: a scale-dependent cascade. Marine Ecology Progress Series, 171, 59–70. Hamilton, L.J., 2001. Acoustic Seabed Classification Systems. Fishermans Bend, Victoria, Australia: Defense Science and Technology Organisation, Aeronautical and Maritime Research Laboratory, DSTO-TN-0401. Haralick, R.M.; Shanmugan, K., and Dinstein, I., 1973. Textural features for image classification. IEEE Transactions on Systems, Man, and Cybernetics, 3(6), 610–621. Hutin, E.; Simard, Y., and Archambault, P., 2005. Acoustic detection of a scallop bed from a single-beam echosounder in the St. Lawrence. ICES Journal of Marine Science, 62, 966–983. Irish, J.L.; McClung, J.K., and Lillycrop, W.J., 2000. Airborne lidar bathymetry: the SHOALS system. PIANC Bulletin., 103, 43–53. Jackson, E.L.; Attrill, M.J., and Jones, M.B., 2006. Habitat characteristics and spatial arrangement affecting the diversity of fish and decapod assemblages of seagrass (Zostera marina) beds around the coast of Jersey (English Channel). Estuarine, Coastal and Shelf Science, 68, 421–432. Jumars, P.A. and Ekman, J.E., 1983. Spatial structure within deepsea benthic communities. In: Rowe, G.T. (ed.), The Sea: Deep-Sea Biology, Volume 8. New York: John-Wiley & Sons, pp. 399–452. Kostylev, V.E.; Courtney, R.C.; Robert, G., and Todd, B.J., 2003. Stock evaluation of giant scallop (Placopecten magellanicus) using highresolution acoustics for seabed mapping. Fisheries Research, 60, 479–492. Legendre, P. and Legendre, L., 1998. Numerical Ecology. Amsterdam: Elsevier, 853p. Legendre, P.; Ellingsen, K.E.; Bjørnbom, E., and Casgrain, P., 2002. Acoustic seabed classification: improved statistical method. Canadian Journal of Fisheries and Aquatic Sciences, 59, 1085–1089. ´ tude hydrodynamique, se´dimenLong, B. and Desrosiers, G., 2006. E tologique et biologique des sites de Maria, Saint-Sime´on, Bonaventure, Newport et Cap-d’espoir dans la baie des Chaleurs, Que´bec, Canada. Ministe`re des Transports du Que´bec, contrat 3100-00QH02. 121 p. + 4 annexes. Lucas, R.; Rowlands, A.; Brown, A.; Keyworth, S., and Bunting, P., 2007. Rule-based classification of multi-temporal satellite imagery for habitat and agricultural land cover mapping. Journal of Photogrammetry and Remote Sensing, 62(3), 165–185. Musick, H.B. and Grove, H.D., 1991. Image texture measures as indices of landscape pattern. In: Turner, M.G. and Gardner, R.H. (ed.), Quantitative Methods in Landscape Ecology, the Analysis and Interpretation of Landscape Heterogeneity. Berlin: Springer-Verlag. Park, J.Y.; Shrestha, R.L.; Carter, W.E., and Tuell, G.H., 2001. Landcover classification using combined ALSM (LiDAR) and color digital photography. ASPRS Conference (St. Louis, Missouri, The American Society for Photogrammetry and Remote Sensing) April 23–27. Pe’eri, S. and Philpot, W., 2007. Increasing the existence of very shallow-water LiDAR measurements using the red-channels waveforms. IEEE Transactions on Geoscience and Remote Sensing, 45(5), 1217–1223. Pope, R.W.; Reed, B.A.; West, G.W., and Lillycrop, W.J., 1997. Use of an airborne laser depth sounding system in a complex shallowwater environment. In: Proceedings of XVth International Hydro Conference (Monaco), p. 10. Research Systems, 2005. IDL User’s Guide, IDL Version 6.2. Boulder, Colorado: Research Systems, Inc. Richards, J.A., 1993. Remote Sensing Digital Image Analysis: An Introduction. Berlin: Springer-Verlag. Riley, J.L., 1995. Evaluating SHOALS bathymetry using NOAA
Journal of Coastal Research, Special Issue No. 62, 2011
98
Collin, Long, and Archambault
hydrographic survey data. In: Proceedings of 24th Joint Meeting of UJNR Sea-Bottom Surveys Panel (Tokyo, Japan), 10 p. Rosenfield, G.H. and Fitzpatrick-Lins, K., 1986. A coefficient of agreement as a measure of thematic classification accuracy. Photogrammetric Engineering and Remote Sensing, 52, 223–227. Ruiliang, P.; Gong, P.; Michishita, R., and Sasagawa, T., 2008. Spectral mixture analysis for mapping abundance of urban surface components from the Terra/ASTER data. Remote Sensing of Environment, 112(3), 939–954. Salovaara, K.J.; Thessler, S.; Malik, R.N., and Tuomisto, H., 2005. Classification of Amazonian primary rain forest vegetation using Landsat ETM+ satellite imagery. Remote Sensing of Environment, 97, 39–51. Siwabessy, P.J.W.; Penrose, J.D.; Fox, D.R., and Kloser, R.J., 2000. Botom classification in the continental shelf: a case study for the North-West and South-West shelf of Australia. In: Proceedings of Australian Acoustical Society Conference (Joondalup, Australia), 6 p. Swain, P.H. and Davis, S.M., 1978. Remote Sensing: The Quantitative Approach. New York: McGraw-Hill.
Syvitski, J.P.M., 1992. Marine geology of Baie des Chaleurs. Ge´ographie physique et quaternaire, 46(3), 331–348. Tuell, G.H., 2002. Data fusion of airborne laser data with passive spectral data. 3rd Annual Airborne Hydrography Workshop (Corte Madiera, California). Turner, M.G.; Gardner, R.H., and O’Neill, R.V., 2001. Landscape Ecology in Theory and Practice: Pattern and Process. New York: Springer Science+Business Media, LLC, 404 p. Underwood, A.J. and Chapman, M.G., 1998. A method for analysing spatial scales of variation in composition of assemblages. Oecologia, 117, 570–578. UNEP (United Nations Environment Programme), 2006. Marine and Coastal Ecosystems and Human Well-Being: A Synthesis Report Based on the Findings of the Millennium Ecosystems Assessment. Nairobi: UNEP, 64p. Wang, C. and Philpot, W., 2007. Using airborne bathymetric lidar to detect bottom type variation in shallow waters. Remote Sensing of Environment, 106, 123–135. Zajac, R.N., 2008. Challenges in marine, soft-sediment benthoscape ecology. Landscape Ecology, 23, 7–18.
Journal of Coastal Research, Special Issue No. 62, 2011
