Digitizing the coral reef: Machine learning of underwater spectral images enables dense taxonomic mapping of benthic habitats

Abstract

Coral reefs are the most biodiverse marine ecosystems, and host a wide range of taxonomic diversity in a complex spatial community structure. Existing coral reef survey methods struggle to accurately capture the taxonomic detail within the complex spatial structure of benthic communities. We propose a workflow to leverage underwater hyperspectral image transects and two machine learning algorithms to produce dense habitat maps of 1150 m2 of reefs across the Curaçao coastline. Our multi‐method workflow labelled all 500+ million pixels with one of 43 classes at taxonomic family, genus or species level for corals, algae, sponges, or to substrate labels such as sediment, turf algae and cyanobacterial mats. With low annotation effort (only 2% of pixels) and no external data, our workflow enables accurate (Fbeta of 87%) survey‐scale mapping, with unprecedented thematic detail and with fine spatial resolution (2.5 cm/pixel). Our assessments of the composition and configuration of the benthic communities of 23 image transects showed high consistency. Digitizing the reef habitat and community structure enables validation and novel analysis of pattern and scale in coral reef ecology. Our dense habitat maps reveal the inadequacies of point sampling methods to accurately describe reef benthic communities.

Crustose coralline algae can contribute more than corals to coral reef carbonate production

“…Additionally, future work could explore the use of microerosion metres, and laser scanning of tidally-exposed reef upper surfaces, as previously used on rock reefs to measure erosion with sub-mm scale precision 113,114 . New methods have also been developed using machine learning that can successfully delineate coralline algae in large scale hyperspectral imagery 115 . These methods all offer promise.…”

Sizing mudsnails: Applying superpixels to scale growth detection under ocean warming

“…Imaging data have proliferated throughout studies of ecology and evolution (Høye et al., 2021; Schürholz & Chennu, 2023; Weinstein, 2018). Digital images are data‐rich, conventionally represented as matrices of pixel intensities across three colour channels (red, green and blue; RGB) with millions of colour variations possible for each pixel in a 24‐bit image.…”

“…Benthic ecosystems are generally data‐poor compared with pelagic zones (Hughes et al., 2021). Benthic habitats classifications with acoustics (Brown et al., 2011; Mehler et al., 2018), scanning images or videos on coral reefs (e.g., Pizarro et al., 2017; Schürholz & Chennu, 2023), or direct observations of epifauna (e.g., Piechaud et al., 2019) all help access species behaviour and interactions, but also omit infaunal groups that possibly comprise the greatest fraction of biodiversity and biomass on Earth (Snelgrove, 1998). Here, extractive sampling (e.g., sediment cores) to monitor specimens under controlled conditions still carries great value for high‐frequency and repeated measurements of individuals.…”

Scalable semantic 3D mapping of coral reefs with deep learning

“…Furthermore, the required equipment should be available worldwide within a reasonable budget. While hyperspectral sensors can facilitate identification of corals (Asner et al., 2020; Chennu et al., 2017; Schürholz & Chennu, 2023), their price can be prohibitive for widespread use. On the other hand, the cost of underwater colour cameras has dramatically fallen, which suggests computer vision tools can be applied to successfully scale automated semantic segmentation of living corals from colour camera imagery.…”

“…While many computer vision tools have been proposed to aid in coral reef monitoring, the scalability in terms of fully analysed transects (in terms of 3D reconstruction and pixel characterization of benthic classes) per time and money unit remains limited. The main lines of computer vision work can be summarized into computer‐aided mapping, commonly realized with structure‐from‐motion (SfM) photogrammetry (Alonso et al., 2019; Bongaerts et al., 2021; Burns et al., 2015; Hopkinson et al., 2020; Leon et al., 2015; Raoult et al., 2017; Storlazzi et al., 2016; Urbina‐Barreto et al., 2021) and benthic cover analysis systems, which use machine learning to recognize coral types and other benthic classes in images (Beijbom et al., 2012, 2015; Chen, Beijbom, et al., 2021; Schürholz & Chennu, 2023; Williams et al., 2019). However, underwater environments pose particular challenges to computer vision methods due to difficult lighting conditions and diffraction effects, caustics, non‐linear attenuation, and scenes with many dynamic objects.…”