Bioacoustics and Machine Learning as Key Tools in Coral Reef Restoration

Le résumé fourni par la source

Coral reef ecosystems have been degraded by a combination of threats at the global scale. In response, local communities, governments, and NGOs are innovating to find ways to manage and restore these vital systems. Rigorous, repeatable, and scalable monitoring metrics are essential for helping managers assess coral reef health and evaluate the effectiveness of conservation and restoration efforts. Traditional monitoring techniques, like fish count transects, are labor intensive and typically capture only brief snapshots of reef life during calm weather and in daylight hours. Passive acoustic monitoring (PAM) is a promising technique that has been successfully applied to expand the spatial and temporal scale of a number of terrestrial and marine mammal monitoring applications. Recent advances in acoustic recording hardware and machine learning techniques have now made it possible for researchers to test passive acoustic surveys as method for monitoring coral reef systems at scale. Here we present results of passive acoustic surveys we conducted in Mo'orea, French Polynesia to explore how PAM could a) expand on traditional monitoring efforts at long-term ecological research sites and b) measure outcomes of a coral outplanting restoration project. We deployed two different types of marine automated recording units (ARUs) from August to November in 2023 at six Mo'orea Coral Reef Long Term Ecological Research (LTER) sites that were either coral or algae dominated ($\mathbf{n} \boldsymbol{=} \mathbf{3}$per treatment). Three ARUs were placed at each LTER sight across reef habitat: fore reef, fringing reef, and back reef. Additional ARUs were placed at a coral outplanting restoration site and a nearby unmanipulated control site. We developed a convolutional neural network model (CNN) to detect and classify target acoustic signals on survey recordings. We present data from four ecologically important signals: ‘parrotfish grazing’ and ‘damselfish calls', a “knock” signal produced by an unknown fish species, and a metric of phonic richness for signals generated by fish. The ‘parrotfish grazing’ class is the scraping sound various species of parrotfish make when scraping algae from the surface of hard coral rock. ‘Damselfish calls’ are vocalizations produced by Pomacentrid fish species. Both parrotfish and damselfish have been shown to be indicators of reef health. Phonic richness, a count of unique fish sounds over a specific time period (in this case a twenty four hour period), has been shown to correlate with reef health and diversity of reef fishes. We present differences in acoustic detection rates between three coral dominated LTER sites and three algal dominated LTER sites for the four metrics. Preliminary results show mean parrotfish grazing detections per minute and fish knock call rates to be lower in the coral dominated LTER's but the difference was not significant ($p =0.18, \mathrm{p}=0.45$respectively). Damselfish calls per minute were higher in the coral dominated LTER's but the difference was not significant ($p=0.71$)). There was no difference in the phonic richness index$(=0.9113)$. We also compared the detection rates at the experimental coral restoration and control sites for the four metrics, but with sample size$\mathbf{n}=\mathbf{1}$, these results are anecdotal. We also found all four metrics to be different between reef types. This study shows the potential for passive acoustic monitoring of coral reef communities, highlights some of the challenges that still need to be addressed to improve on species detection and classification, and provides practical examples of ways to efficiently measure outcomes of coral reef restoration efforts.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.