Research papersUnsupervised fuzzy classification and object-based image analysis of multibeam data to map deep water substrates, Cook Strait, New Zealand
Highlights
►Multibeam survey of Cook Strait was used to generate a regional substrate classification map. ► Ranges of water depths, seafloor substrates and geological landforms were classified. ► Substrate classes are identified using an unsupervised classification technique. ► Texture analysis identified classes consistent with backscatter angular responses for sediments. ► Object-based image analysis allows for correlations to be made between physical and acoustic data.
Introduction
Quantifiable methods that map benthic substrates from remotely sensed data have become a significant focus of the seafloor mapping community (e.g., Brown and Blondel, 2009, Le Bas and Huvenne, 2009, Lucieer and Lucieer, 2009, Wright and Heyman, 2008). Seafloor sediments and geomorphology are valuable proxies for: (i) substrate composition and sedimentary processes; (ii) tectonic processes; (iii) benthic habitat (Brown and Blondel, 2009, Kostylev et al., 2001, Orpin and Kostylev, 2006, Ryan et al., 2007) and (iv) benthic biodiversity (Schlacher et al., 2009, Thrush et al., 2006). Moreover, knowledge of the distribution of these proxies informs species management and the development of marine protected areas (Anderson et al., 2008, Anderson et al., 2007, Jordan et al., 2005, Kostylev et al., 2003, Lucieer and Pederson, 2008).
Acoustic backscatter images generated from multibeam echo sounder (MBES) data are often complex due to variability in seafloor roughness and impedance, sediment grain size and volume heterogeneity, as well as the inherently noisy signal at nadir and steep grazing angles (Marsh and Brown, 2009). For these reasons, pixel-based classification schemes are critically limited when differentiating seabed zones with well-defined boundaries, resulting in inconsistent classifications that are highly biased by the specular response at nadir. In contrast, object-based image analysis techniques overcome these difficulties by first segmenting the image into meaningful seabed zones of various sizes, based on their spectral and spatial characteristics (Blaschke, 2010). Such classification methods, originally developed for remotely sensed data from satellites, have been successfully applied to multibeam and sidescan sonar data (Lucieer, 2007, Lucieer, 2008, Preston, 2009).
In this paper we apply an object-based image analysis technique to segment and classify multibeam backscatter data from Cook Strait, New Zealand, to generate a first-level quantitative seabed substrate map. The diverse geological, hydrodynamic and ecological environments of Cook Strait, together with comprehensive multibeam bathymetric (Fig. 1) and backscatter coverages (Fig. 2), alongside a wealth of geological samples, provide an excellent opportunity to test and develop automated methods to characterise marine habitats over large areas using the image analysis approach.
Central to our study is the feasibility and validity of applying an unsupervised method over a large and geomorphologically complex area that contains diverse natural environments. This is achieved through two key steps: (i) unsupervised fuzzy-c-means (FCM) classification of sediment samples to identify the appropriate number of classes and their core spatial areas and (ii) use of these areas as training samples in an object-based image analysis of MBES imagery to derive a geomorphological map for the whole study area. Finally, we present a validation of our results.
In the first stage, we apply a quantitative technique employing an unsupervised FCM classification that identifies the optimal number and spatial location of seabed classes, based on seafloor sediment samples (Lucieer and Lucieer, 2009). The methods employed here show that FCM can locate potential boundaries and transition zones between multivariate seabed properties, which correlate strongly with backscatter, thus providing a potentially powerful tool to quantify uncertainty associated with transition zones between the classes. The degree of success of unsupervised classification is notoriously difficult to quantify (Canty, 2007, Duda and Canty, 2002) because there is rarely a reference information to assess the accuracy of the classification results or there is a lack of independent data to validate the results. In this paper, we validate the clustering result by comparing the results with sediment and geomorphological maps of Cook Strait.
In the second stage, the acoustic backscatter image is segmented based on texture, and grey scale values and the size and shape of these segments are differentiated by their depth. An object-based image analysis framework is employed to classify image segments into geomorphological classes based on stage one. Traditionally, benthic habitat classes are assigned based on expert knowledge of the seafloor or visual discrimination of particular textures. In this study, we propose a more automated and objective approach towards seabed classification.
The diverse and complex geomorphology of Cook Strait is the result of the dynamic climatic, tectonic and hydrodynamic forcings that have impacted on the region since at least the last post-glacial period (Lewis et al., 1994, Proctor and Carter, 1988). The 20–60-km-wide oceanic Strait separates the North and the South Islands of New Zealand. Extremely vigorous winds are driven by the prevailing westerly airflow of the Tasman Sea (Fig. 1) across mountainous New Zealand, and steered and intensified through Cook Strait (Harris, 1990). As a result, winds are strong and distinctly bimodal: prevailing winds are from the south and from the north to northwest. High winds result in an equally vigorous wave climate, and when motions are superimposed on the Strait's powerful tidal currents, the result is a highly dynamic marine environment at all water depths. The lunar semi diurnal (M2) tide is a trapped wave travelling anti clockwise around New Zealand. High tide on the west coast occurs close to low tide on the east coast, the Strait effectively short circuiting the phase difference resulting in rapid tidal currents. These powerful currents coupled with strong sediment delivery, marked effects of eustatic changes in sea level, strong meteorological forcing of the circulation, and active tectonism (Hicks and Shankar, 2003, Lewis et al., 1994) have produced distinctive patterns of sedimentation, which include sediment wave fields (Carter, 1992, Lamarche et al., 2010), coarse gravelly shelf, mud flats, deep-sea canyons (Mountjoy et al., 2009) and deep turbidite filled channels (Lewis et al., 1994).
The region lies immediately west of the southern termination of the active convergent Pacific–Australia plate boundary, which generates intense seismic activity and associated ground-shaking in the region (Pondard and Barnes, 2010, Stirling et al., 2000). The continental shelf and the slope are dissected by active faults and NE-trending structurally controlled ridges that are the locus for pronounced slope instability (Barnes and Audru, 1999). Semi-circular slump scarps are up to 1100 m in height with evidence of ample debris at their base and in channel axes (Mountjoy et al., 2009).
The shelf break ranges from ∼50 m water depth in the Narrows to ∼150 m in the eastern Cook Strait. Eastward, beyond the continental slope, the homogeneously flat-floored Hikurangi Trough contains a deep turbidite fill (Lewis et al., 1994). During the Last Glacial Maximum, the continental shelf was emergent as a coastal plain (Carter, 1992, Lewis et al., 1994), most of which was subsequently blanketed by a wedge of post-last glacial sediment. Today, some areas of the shelf remain bare of post-last glacial mud that the erosional surface outcrops at the seabed (Mountjoy et al., 2009). Evidence of active and relict fluid flow and seafloor seeps, usually associated with dense concentrations of distinctive, chemosynthetic biota have been reported in the SE approaches of Cook Strait (Barnes and Audru, 1999, Law et al., 2009).
Section snippets
Backscatter and bathymetry datasets
MBES data were collected during six oceanographic surveys of R.V. Tangaroa between 2001 and 2005 in Cook Strait by the National Institute of Water and Atmosphere Research (NIWA) (Lamarche et al., 2010). The total surveyed area is ca. 8500 km2 at depths >100 m (Fig. 1). R.V. Tangaroa has a hull-mounted Kongsberg EM300 (32 kHz) MBES, that fully compensates the vessel position and motion (heave, pitch, roll and yaw). Sound velocity was measured at regular intervals to account for hydrological
FCM methods
The FCM clustering algorithm (also known as the fuzzy-k-means algorithm) is an unsupervised classification algorithm designed to identify groups of samples sharing similar characteristics in a multivariate feature space (Bezdek et al., 1984, Dunn, 1973). Fuzzy clustering provides a means of spatially depicting classification uncertainty related to transition zones and class overlap, which is not possible with the non-spatial statistics derived from other accuracy assessments such as error
OBIA methods
Object-based image analysis (OBIA) is a relatively new image processing technique that classifies objects in an image, based on a segmentation phase followed by an object-based classification phase (Benz et al. 2004). A segmentation algorithm was used to subdivide the combined multibeam backscatter and bathymetric images into smaller image objects based on an optimisation procedure, which locally minimises the average heterogeneity of image objects for a given resolution across the image. Image
Classification results and discussion
The feature space optimisation method in Definiens Developer© (version 8.0) was employed to identify the best combination of textural and object features most suitable for separating the classes based on the selected samples from the FCM, in conjunction with the nearest-neighbour classifier. The best separation distance that we achieved between the two segmentations was 1.18 for Segmentation 50 and 1.14 for Segmentation 100, which indicated that the fine-scale segmentation level provided a
Conclusion
We used an object-based segmentation of the backscatter and bathymetry data to generate a four-class classification map from the geomorphologically diverse Cook Strait, New Zealand. This quantitative first order correlation between the seabed substrate and the MBES data is a first for Cook Strait at a regional scale. We use the backscatter data as a proxy for the seafloor sediment grain size and heterogeneity. The four-class classification map shows some correlation with the sediment map of the
Acknowledgments
Anne-Laure Verdier (NIWA) processed the backscatter data using the Sonarscope software of Ifremer. Arko Lucieer (UTAS) and Alan Orpin (NIWA) provided very valuable comments on this manuscript. This project was funded by the Royal Society of New Zealand, Bilateral Research Action Program of the International Science and Technology Fund, the New Zealand Foundation for Research Science and Technology (FRST, programme C01X0702) and the Commonwealth Environment Research Facilities (CERF) program, an
References (49)
- et al.
Tectonic and geological framework for gas hydrates and cold seeps on the Hikurangi subduction margin, New Zealand
Mar. Geol.
(2010) - et al.
Multi-resolution, object-oriented fuzzy analysis of remote sensing data for GIS-ready information
ISPRS J. Photogramm.
(2004) - et al.
FCM: the fuzzy c-means clustering algorithm
Comput. Geosci.
(1984) Object-based image analysis for remote sensing
ISPRS J. Photogramm.
(2010)- et al.
Fine-scale sedimentary structure: implications for acoustic remote sensing
Mar. Geol.
(2002) - et al.
Developments in the application of multibeam sonar backscatter for seafloor habitat mapping
Appl. Acoust.
(2009) - et al.
Angular range analysis of acoustic themes from Stanton Banks Ireland: a link between visual interpretation and multibeam echosounder angular signatures
Appl. Acoust.
(2009) - et al.
Stock evaluation of giant scallop (Placopecten magellanicus) using high-resolution acoustics for seabed mapping
Fish. Res.
(2003) - et al.
Acquisition and processing of backscatter data for habitat mapping—comparison of multibeam and sidescan systems
Appl. Acoust.
(2009) - et al.
The opening of Cook Strait: interglacial tidal scour and aligning basins at a subduction to transform plate edge
Mar. Geol.
(1994)