Using object-based image analysis to determine seafloor fine-scale features and complexity

Abstract

Autonomous and remotely operated underwater vehicles equipped with high-definition video and photographic cameras are used to perform benthic surveys. These devices record fine-scale (< 1 m) seafloor features (seafloor complexity) and their local (10-100s m) variability (seafloor heterogeneity). Here, we introduce a methodology to efficiently process this optical imagery using object-based image analysis, which reduces the pixels in high-resolution digital images into a collection of "image-objects" of homogeneous color and/or luminosity. This approach uses intuitive user-defined parameters and reproducible computer code, which aims to facilitate comparisons between habitats and geographic regions. We test this methodology with 511 images taken on the seafloor of a glaciated continental shelf (Gulf of Maine, northwest Atlantic), and describe three applications: (1) estimating percent cover of conspicuous epibenthic fauna by building a Random Forest binary classifier assigning an identity to image-objects; (2) correlating image complexity (number of image-objects) with mean particle grain size; and (3) estimating seafloor heterogeneity from local variability in image complexity within and between two physiographic regions. Percent cover of epibenthic fauna estimated by the Random Forest binary classifier was in close agreement with the human visual assessment. Mean particle grain size (/ scale) was inversely correlated with image complexity (maximum Spearman's q 5 20.89, p < 0.01) with images dominated by pebbles, cobbles, boulders (low on / scale) yielding high image complexity. Predictive relationships of sediment composition were established using polynomial regression. Lastly, our approach could differentiate habitats within and between physiographic regions by using mean seafloor complexity and local variability along transects.Benthic habitat structure is composed of two factors: complexity, the absolute abundance of structural components (e.g., rocks, mounds); and heterogeneity, the variation in complexity due to changes in the relative abundance of these structural components (McCoy and Bell 1991;Sebens 1991). On sedimented substratum, complexity arises from the distribution of sediment grain size and small-scale topographic features (e.g., pits and burrows) resulting from local hydrodynamics and bioturbation. On hard substratum, in rocky intertidal and subtidal habitats, complexity is most often measured with a surface to area ratio [e.g., fractal dimensions (Kostylev et al. 2005), and the chain-and-tape approach (Risk 1972)], which is influenced by the presence of crevices, rock walls or biological structures. Seafloor complexity influences larval settlement (e.g., Walters and Wethey 1996), the abundance and diversity of invertebrate assemblages (e.g., Kostylev et al. 2005;Matias et al. 2010), predator-prey interactions (e.g., Grabowski 2004, and the concentration of organic matter (e.g., Abelson et al. 1993).Among other factors, spatial variability in fine-scale (< 1 m) seafloor complexity (i.e., sea...

Comparing the Performance of a Remotely Operated Vehicle, a Drop Camera, and a Trawl in Capturing Deep-Sea Epifaunal Abundance and Diversity

“…Lastly, imagery appears to have greater catchability for sessile fauna compared to trawl, yet is a more timeintensive to process. Thus, research into automation of image processing, will also be a great benefit to future deep-sea research (e.g., Lacharité et al, 2015).…”

A Systematic Review of Remotely Operated Vehicle Surveys for Visually Assessing Fish Assemblages

“…High-definition video cameras carried as extra payload provide researchers with permanent records of biota and their habitat associations (Macreadie et al, 2018). While some studies used photography to makes these assessments, these studies were typically focused on mega-benthic taxa (Salvati et al, 2010;Thresher et al, 2014;Lacharité et al, 2015;Cánovas-Molina et al, 2016), and may not be ideal for moving targets such as fish.…”

Learning deep-sea substrate types with visual topic models

“…[9] proposes a method to measure seabed complexity by segmentation, and applies a random forest classifier on those segments to identify certain objects in a supervised manner. [16] develops a method based on Self Organizing Maps (SoM) to learn a feature representation for segmenting seabed images, and shows that they successfully identify metallic nodules with a simple supervised classifier in the learned feature space.…”