An automatic and quantitative approach to the detection and tracking of acoustic scattering layers

Acoustic scattering layers are ubiquitous, horizontally extensive aggregations of both vertebrate and invertebrate organisms that play key roles in oceanic ecosystems. However, currently there are no conventions or widely adaptable automatic methods for identifying these often dynamic, spatially complex features, so it is difficult to consistently and efficiently describe and compare results. We developed an automatic scattering layer detection method that can be used to monitor changes in layer depth, width, and internal structure over time. Extensive, contiguous regions of the water column that have echo strengths above a threshold were identified as “background layers.” They correspond to regions of the water column that contain scattering from diffusely distributed organisms. Often, background layers contained contiguous, horizontally extensive features of concentrated acoustic scattering we identified as “strata.” These features were identified by fitting Gaussian curves to the echo envelope of each vertical profile of scattering, and their boundaries were identified as the endpoints of the region containing 95% of the area under the fitted curves. These endpoints were linked horizontally to make continuous tracks. Bottom and top tracks were paired to identify features that sometimes extended horizontally for tens of kilometers. This approach was effective in three disparate ecosystems (the Gulf of California, Monterey Bay, and the Bering Sea), and a sensitivity analysis showed its robustness to changes in input parameters. By allowing a comparable, automated approach to be used across environments, this method promotes the improved classification and characterization of acoustic scattering layers necessary for examining their role in oceanic ecosystems. *Corresponding author: E-mail: [email protected] Acknowledgments The authors wish to thank Chad Waluk for technical assistance, Sarah Emerson for consultations on statistics, and Marisa Litz, Emily Shroyer, and Scott Heppell for helpful comments. We thank the US Office of Naval Research (N00014-11-1-0146) for dedicated support of the data analysis and the National Science Foundation (0851239), North Pacific Research Board (Bering Sea Projects B67 and B77), and the US Office of Naval Research (N0014-05-1-0608) for supporting the collection of the data used. DOI 10.4319/lom.2014.12.742 Limnol. Oceanogr.: Methods 12, 2014, 742–756 © 2014, by the American Society of Limnology and Oceanography, Inc. LIMNOLOGY and OCEANOGRAPHY: METHODS for identifying layers or even agreed upon conventions for describing basic features of scattering layers (i.e., determining their boundaries, characterizing their acoustic structure, and describing their depth). While some authors measure the depth of a layer in the water column from the top boundary (Baliño and Aksnes 1993; Tont 1976), others focus on the bottom (Kumar et al. 2005), and others on the depth of peak energy (Benoit-Bird et al. 2010). However, layers can have variable internal structure and can be more than 100 m thick, so considering only one measure of depth within the water column may be insufficient for studying the response of scattering layers to oceanographic parameters. Additionally, describing acoustic structure within scattering layers has had only limited emphasis in the literature (e.g., Benoit-Bird and Au 2003) despite evidence that this structure can have important implications for the ecology of scattering layer organisms (Benoit-Bird and McManus 2012). Our goal was to create a method for automatic layer detection that could assist in addressing these research gaps by consistently quantifying the spatial characteristics of scattering layers. A number of approaches have been tried previously but have met with mixed success. The most common approaches have involved basic visual examination of echograms (e.g., Kumar et al. 2005; Robinson and Gómez-Gutiérrez 1998; Simard and Mackas 1989). This approach is simple, can be done in real-time for determining appropriate net trawl depths, and can be used to effectively determine layer boundaries even in acoustically complex environments; however, as with any procedure done by trained observers, results can vary between individuals as well as for one individual observer over time (as discussed in Jech and Michaels 2006). Visual identification is also laborious and can prove intractable due to the substantial volume of data that can be generated by acoustic instruments on ships and other platforms. Another common approach has been to classify layers based on the broad depth bin in which acoustic energy is contained (i.e., surface, mesopelagic, epipelagic, and bathypelagic). This approach allows researchers to ignore the structure of layers and to focus solely on the acoustic energy that is present at a given depth of interest. While useful for studies of DVM and for wide area biomass surveys (e.g., Kloser et al. 2009), this approach does not allow individual features to be tracked, nor does it exclude acoustic energy that is not part of a scattering layer. Automatic layer detection approaches have the potential to overcome the drawbacks of manual layer identification and depth-based layer definitions. To do so effectively they must locate the top and bottom of several layers in the water column, identify layers with a variety of acoustic structures, be effective in a range of locations with differing biological layer compositions, respond predictably to changes in input parameters, and account for horizontally extensive layers that can change their depth in the water column and can split and merge over time. A typical automated approach looks for the sharpest gradient in the water column. Often used for seafloor detection, such approaches can be adept at locating the depth of long, continuous features. These types of algorithms, however, locate only a single depth and thus ignore layer thickness. It can also be difficult for seafloor detection methods to consistently locate biological layers since, unlike the seafloor that exhibits a sharp gradient in scattering relative to open water, scattering from biology typically diminishes from the peak more slowly and is only rarely characterized by a sudden increase or decrease in echo strength. School detection algorithms (e.g., Barange 1994) have been adapted to identify scattering layers since they are designed to find the boundaries of discrete aggregations of organisms. Similar to many bottom detection algorithms, they work by looking for an amplitude difference between a region and its surrounding regions, typically searching for a value greater than a fixed threshold. As shown by Burgos and Horne (2007), however, the choice of acoustic threshold has a significant effect on the height, length, depth, and the total acoustic energy of the detected aggregations, so choosing an appropriate, robust acoustic threshold can be challenging. Weber et al. (2009) used a statistical approach to determine an appropriate noise-threshold for analysis of fish schools, relying on a controlled situation with data collected in the presence and absence of a single-species assemblage. Scattering layers, in contrast, pose a particularly challenging scenario since they are characterized by mixed species aggregations, are horizontally extensive, and often exist in regions with varying background conditions. Perhaps the most significant challenge for adapting school and patch detection algorithms to layers is the requirement that identified features be limited in their horizontal extent. Nero and Magnuson’s (1992) patch detection algorithm, for example, limits the spatial extent of patches to the size of their chosen smoothing window, whereas Weill et al.’s algorithm (1993) must prematurely terminate long features. School detection algorithms also assume that features are relatively stable, yet scattering layers are defined partly by their evolution over time. Fig.!1, for example, features a deeper layer shoaling at sunset to meet a shallow layer. A school detection algorithm would treat these two features as a single combined aggregation despite clear differences in their behavior. To account for properties unique to scattering layers, some novel approaches have been developed. Bertrand et al. (2010) successfully tracked the bottom of a layer of pelagic organisms by determining the depth at which 98% of accumulated echoes occurred. In their study area, this depth was shallow (<100 m) and characterized by a rapid build-up of echo energy at the base of the layer due to a strong oxycline. However, this approach is not effective for layers that have more graded changes in scattering or in regions containing several layers that distribute echo energy throughout the water column. Benoit-Bird and Au (2004) identified scattering layer boundaries based on differences in the numerical density of organCade and Benoit-Bird Detection of acoustic scattering layers