Coral reefs are among the planet’s most vital and diverse ecosystems, yet they are increasingly under threat from human activity and climate change. Passive Acoustic Monitoring offers a scalable, non-invasive way to assess reef health by listening in on their underwater soundscapes. Healthy reefs burst with the crackle of snapping shrimp, melodic fish calls, and a vibrant chorus of marine life, whereas degraded reefs fall silent and eerily quiet. This study explores whether the SurfPerch foundation model, originally trained on tropical reef audio, can accurately distinguish between healthy and degraded coral reefs in the sub-tropical, turbid waters of Moreton Bay, Australia.
The Soundscape
Healthy coral reefs produce a dense, continuous acoustic signature dominated by snapping shrimp clicks and fish vocalizations in the 512–2048 Hz range. Degraded reefs, stripped of their biological communities, are sparse and transient-heavy. The difference is striking even as a spectrogram.
Healthy, Myora Reef
Degraded, Goat Island
Mel spectrograms of 5-second clips from each site. Myora shows sustained broadband biophony; Goat Island is acoustically impoverished.
Study Sites
Both sites are sub-tropical reefs in Moreton Bay, Queensland: high-latitude, turbid, and ecologically marginal. Myora Reef is flushed by oceanic water via Rainbow Channel and supports 23–42% Acropora coral cover. Goat Island, just a few kilometers away, is sediment-impacted from riverine runoff and still recovering from a 2022 flood event.
Healthy reference
Degraded, sediment-impacted
Moreton Bay, Queensland. Recorded by us, over four days in November 2025 using a Cetacean Research CRT-40P hydrophone at 48 kHz / 24-bit.
Method
80 min of underwater recordings
Artifact removal, 32kHz mono
7,973 labeled windows
1,280-dim embeddings (frozen)
Linear + neural network sweep
Healthy or Degraded
Embedding Space
Projecting the 1,280-dimensional embeddings into 2D with t-SNE reveals clear structure. Healthy Myora samples form a dense, coherent cluster, while degraded Goat Island samples are more dispersed. The partial overlap suggests some acoustic similarity at the margins, but the bulk of the data separates cleanly.
t-SNE projection of 7,973 five-second audio segments. Green: Healthy (Myora Reef). Purple: Degraded (Goat Island).
Results
SurfPerch embeddings dramatically outperform traditional MFCCs across every metric. The model achieves 0.94 AUC with a simple linear head on frozen embeddings, no fine-tuning required. The Gini coefficient nearly doubles (0.89 vs 0.50), indicating far stronger discriminative power.
ROC curves. SurfPerch (AUC = 0.94) vs MFCC baseline (AUC = 0.75). Optimal threshold at 0.82.
Confusion matrix at threshold 0.82. Balanced recall: 89.6% degraded, 91.6% healthy.
Gini coefficient. SurfPerch: 0.89. MFCC: 0.50.
