2.1 Datasets

For this study, we pooled three tropical coral reef PAM datasets, each with different supporting ecological or geographical data. These were collected from three distinct coral reef biogeographic realms: i) South Sulawesi, Indonesia, in the Tropical Indo-Pacific; ii) Lizard Island, Australia, in the Coral Sea; and iii) French Polynesia, in the Mid-South Tropical Pacific (Figs 1B and S1) [32]. Recording periods from the Australian and French Polynesian sites were sub-sampled to reduce dataset size (S1 Text). From each dataset, the supporting metadata enabled the division of recording sites into two classes. These classes represented divisions between classes for which the literature suggests there is strong evidence that differences in the soundscape should be expected [5,9,33]. This strategy enabled us to devise tasks with which we could compare the different ML methods in question, but we do not test more challenging differences or gradients at this stage. For the Indonesian data, sites were divided into high (91.2-93.1%) and low (2.1-17.6%) coral cover groups. The Australian sites were divided into two classes, one with high biomass (20.5-52.2 kg) and species richness (45spp - 53spp) scores per transect, and another with low biomass (5.2-9.8 kg) and species richness (30spp – 35spp) scores per transect (S2 Fig), which we refer to as high- and low-fish-diversity sites from here on. The French Polynesian sites were divided into shallow (10-15m) and mesophotic (55-65m) depth classes.

Each dataset was collected using SoundTrap hydrophone recorders (SoundTrap 300ST, Ocean Instruments, Auckland, NZ) which were calibrated by the manufacturer with a flat frequency response. These were suspended 0.5 m above the seabed using a sub-surface buoy. To mitigate against instrument bias, the individual recorders were frequently rotated between the different respective sites within the Indonesian and Australian datasets, however, logistical constraints of deploying recorders at mesophotic depths prevented this for the French Polynesia dataset. The full sampling regime and recorder rotation is detailed in S1 Text. To minimise the introduction of geophonic noise, all recordings were taken during seastates between 0–2 on the Beaufort scale and never during periods forecasted to rain, where conditions deviated from this, recordings were not taken. All recordings were taken in remote areas away from frequently trafficked boating channels and mooring points to minimize the presence of anthropogenic noise, a check for boat noise was also performed to ensure its prevalence was low (S2 Text). In total, 3335, 8127 and 8925 one-minute recordings were used from the Indonesian, Australian and French Polynesian datasets respectively. Further details including, maps, recording schedules, and instrument rotations, are available in S1 Text and S1 Fig.

2.2 Extracting feature embeddings

Each dataset was divided into one-minute recording periods and embeddings were extracted from these. Guidance from previous literature was used to assemble a compound index (Table 1) [15,20,26]. The aim was to include a broad selection of features, whilst preventing this from including unnecessary noise which can reduce the performance of machine learning algorithms. This began with the identification of suitable acoustic indices using the following criteria: (i) previous use in the literature on marine soundscape recordings, (ii) do not require recorder calibration, which would preclude the use of more widely accessible recorders [34]; (iii) could be computed using existing toolkits in a single programming language (e.g., Python). Of the identified indices, the acoustic diversity index (ADI) and acoustic evenness (AEI) have previously been reported to strongly covary [20]. Therefore, to reduce the introduction of noise only AEI was used (Table 1). Seven of the eight indices were then calculated across three different frequency bands: a low-frequency (0.05–2 kHz) where fish sounds dominate [5], a medium-frequency band (2–8 kHz) where snapping shrimp sound dominates [20] and a full band (spanning 0.05–8 kHz). We excluded frequencies below 0.05 kHz from the low- and broad-frequency band recordings to remove geophonic noise and self-noise from the recording system [35]. The exception was the normalised difference soundscape index (NDSI), which requires the input of two bands, for which 0.05–1 kHz and 2–5 kHz bands were used as in Williams et al., (2022) [20]. Recordings were split into 0.96-second segments (to match the CNNs input windows), totalling 62 per minute, and indices were calculated for each segment. The mean and standard deviation of these were then taken for each minute [15], providing a 44-dimension embedding which was used as the compound index to represent each one-minute recording. All processing was performed in Python (v3.7) with the scikit-maad package (v1.3) [36].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Acoustic indices used in the compound index. Indices were calculated using the Scikit-maad (v1.3) package. Where additional settings are ‘None’, the index was calculated over the precomputed spectrogram with no further parameters required. The reference column cites a study which has reported a relationship between the respective index and at least one aspect of coral reef ecology.

https://doi.org/10.1371/journal.pcbi.1013029.t001

For the pre-trained convolutional neural network (P-CNN), we selected VGGish due to its successful application in previous work on terrestrial soundscape data [29,38]. VGGish was trained on the YouTube-100M dataset, a diverse collection of YouTube audio clips totalling 5.2 million hours in length, to produce a highly generalisable audio embedding extractor [31]. It uses a version of the Visual Geometry Group object recognition architecture that was adapted for audio input [39]. The P-CNN pre-processing down-samples recordings to 16 kHz and divides them into non-overlapping 0.96 second audio frames, processed through a Short-time Fourier Transform with 25 ms windows every 10 ms. These are integrated into 64 mel-spaced frequency bins which are log-transformed to produce a 96 × 64 bin log-mel spectrogram which matches the network’s input shape. Further details are available in Hershey et al., (2017) [31]. This P-CNN can be configured to output the 128-dimension feature set from the penultimate layer in place of the classification head. We averaged these feature values from each one-minute recording to produce a single embedding. Processing was performed in Python (v3.7).

To ensure comparability with the P-CNN, for the trained CNN (T-CNN) we trained VGGish on each respective reef soundscape dataset and task from scratch on an NVIDIA A100 GPU. The pre-processing protocol used for the P-CNN was replicated to produce 960 ms log-mel spectrogram samples. The order of these samples was shuffled and minibatches of 32 samples were used during training. The number of output nodes of the network was adjusted to match the number of target classes in each task. The T-CNN was then trained for 50 epochs, with inference on the validation data every epoch. The epoch which reported the highest validation accuracy was then used for inference on the test data. All other hyperparameters followed the default settings for VGGish used in Hershey et al (2017) [31], including randomly initialized weights using a standard deviation of 0.01, a learning rate of 0.001, and an Adam optimiser. Other parameters specific to unsupervised and supervised learning are outlined in sections 2.3 and 2.4 respectively, all other parameters can be found in the vggish_params.py script (see the Code reporting section). Processing was performed using Tensorflow (v1.15) and scikit-learn (v0.22) in Python (v3.7).

2.3 Unsupervised clustering

Unsupervised clustering was used to reveal structures and patterns in the data without relying on pre-labelled classes. Embeddings from the compound index and P-CNN were extracted from recordings. To generate embeddings using the T-CNNs, the VGGish CNN was trained for 50 epochs on each dataset to predict which site recordings originated from within the respective dataset (see section 2.4 for more detail). The final layer of the trained models were then removed to produce three networks comparable to the P-CNN, except these were now trained on reef soundscape recordings from the three respective datasets. These T-CNNs were then used as pretrained embedding extractors on all recordings from their respective datasets in the same manner as the P-CNN. Importantly, only the T-CNN’s trained on the site classes, not habitat classes, were used for embedding extraction in unsupervised learning. This was to avoid introducing ecological context into the model training, which may not be available in exploratory analyses this experiment was simulating, whereas the site of origin should typically be available.

To determine which of the three embedding extraction methods was most proficient at producing embeddings that represented known properties of the data, a qualitative assessment was first performed using UMAP visualisation in two dimensions with UMAP’s associated Python package (v0.5.3) [32]. For all six tasks performed by each method, a plot was produced where points were labelled with their known class, and the fidelity of groupings to their true class compared between each method.

A quantitative assessment was also undertaken using affinity propagation clustering [40,41]. This algorithm was selected as it does not require the number of clusters to be predefined, instead the affinity propagation algorithm identifies this, simulating analysis with an entirely unlabelled dataset where number of classes is unknown. To improve clustering, UMAP was used to reduce the embeddings of each dataset to 10 dimensions [42]. For a given method and task, the cluster to which each recording was assigned was entered into a contingency table against its known class. Models most proficient at clustering recordings from the same class, while excluding recordings from other classes, generated clusters with a higher fidelity to individual classes. Those less proficient assigned recordings more randomly across clusters. A chi-squared test was performed on contingency tables to assess this, where a higher score indicating models with a higher fidelity to true classes. Processing was performed using scikit-learn (v0.22) in Python (v3.7).

2.4 Supervised classifiers

All three ML methods were set two classification tasks, habitat type and individual site identification, from each of the three soundscape datasets. The habitat classification tasks were as follows: high or low coral cover sites for the Indonesian dataset, high- or low-fish-diversity sites for the Australian dataset, and shallow- or mesophotic-depth sites for the French Polynesian dataset. For each task, recordings were split into training, validation and test sets, where 66–75% of the data was used for training and the remaining data was evenly split into validation and test sets (S1 Text). To account for temporal autocorrelation in data and recorder bias, these divisions were carefully performed to exclude entire sites from the training data where possible, or extended contiguous recording periods if not, see S1 Text for full details on each dataset. The validation and test accuracies were carefully monitored to ensure clear over or underfitting did not occur, these values can be found in the cnn_predictions.zip within the supporting Zenodo repository (see Data Availability).

For the compound index and P-CNN embeddings, random forest ML classifiers were trained for each task. Random forests were selected due to their previous successful implementation in the soundscape literature [15,26], robustness to overfitting, and low computational costs, enabling computation on personal devices. For each task, fifty random forest classifiers were trained and then used for inference on the validation data. The instance which reported the highest validation accuracy was used for inference on the test data, and the accuracy of this reported. As classes were well balanced within each task, raw accuracy (the proportion of one-minute recordings correctly classified) was used as the performance metric. For each task, this process was repeated multiple times using carefully selected splits of the data into training, validation and test sets. To ensure the models were tested on novel data that was out-of-distribution from the training data, the data was grouped into blocks of entire sites and/or dates and splits were devised using these to hold out entire blocks for testing, rather than random assignment (S1 Text). To ensure the data splits selected did not bias any approach, one hundred repetitions were performed for each task, except for the Australian and French Polynesian habitat classification tasks, where only 32 were used as this was the maximum number of possible combinations which enabled the exclusion of entire sites from the training data (S1 Text). Analysis was performed using the default random forest parameters in scikit-learn (v0.22) with Python (v3.7).

For the T-CNNs, networks were trained on the reef soundscape dataset for each individual task. The most common class prediction across all 0.96 second segments in a one-minute recording was used as the class prediction for each minute in the test data. This process was repeated 100 times using the same training, validation and test combinations as the compound index and P-CNN. This workflow was performed using Tensorflow (v1.15) in Python (v3.7).

Analysis of variance (ANOVA) tests were used to determine whether significant differences existed between the accuracy of each embedding approach across all repetitions for each task. Where significant differences were detected (p < 0.05), post hoc Tukey tests were used to determine which methods differed; performed using the scipy library (v1.73) in Python (v3.7).

3.1 Exploring coral reef soundscapes with unsupervised learning

Uniform Manifold Approximation and Projection (UMAP) plots were used to quantitatively explore patterns in the soundscape. When interpreting UMAP plots, it is important to note that while UMAP accurately maps local similarities within the data, allowing it to identify clusters, the spatial arrangement of these clusters is less reliable [43]. A visualisation of the three full datasets (Indonesia, Australia, French Polynesia) using the P-CNN UMAP embeddings showed that each dataset formed distinct groups, separate from the others, indicating that the recordings from each location had unique properties (Fig 2A).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. (A) Uniform manifold approximation (UMAP) dimensionality reduction visualisations of all three datasets produced using the pretrained CNN, individual points represent a one-minute recording.

(B) Selected UMAP visualisations of the compound index, pretrained CNN and trained CNN embeddings. visualisations are labelled with colours corresponding to the habitat class (high or low fish diversity) or site. The remaining visualisations can be found in S3 Fig. (C) UMAP visualisation of the French Polynesian dataset which reveals temporal patterns within the dataset produced using the pretrained CNN. The timescale runs from 00:00am (hour of day = 0) to 23:59pm (hour of day = 24).

https://doi.org/10.1371/journal.pcbi.1013029.g002

Qualitative inspection of UMAP visualisations from each individual dataset also revealed that the known habitat classes were frequently key drivers behind the clusters that were formed, indicating the influence of these on the soundscape (Figs 2B, and S3A,S3B and S3C) [29]. Of the three embedding methods, the recordings were most clearly separated into discrete clusters that conformed to habitat classes by the T-CNN, followed by the P-CNN (Figs 2B and S3). The UMAP plots also revealed a unique soundscape “fingerprint” for many sites, though this effect was strongest for the French Polynesian dataset where recording instruments could not be rotated between sites (Figs 2B and S3 S1 Text). In addition to identifying site and habitat classes, UMAP plots also revealed temporal patterns within the data (Fig 2C and S4-6;S3 Text). For example, crepuscular periods linked separate night and day clusters for the French Polynesian dataset (Fig 2C), while the soundscape of two sites in Indonesia converged around the new moon (4-S6 Figs; S3 Text).

Chi-square scores from affinity propagation clustering were highly significant (p < 0.001) for all three embedding types across each dataset (Table 2), indicating the clustering algorithm was able to non-randomly assign recorders to clusters that corresponded to habitat or site classes in every case. Overall, the T-CNNs yielded the highest chi-square scores, indicating greater fidelity to both habitat and site classes. In every instance, this was followed by the P-CNN and then compound index (Table 2). However, for the French Polynesian habitat and site identification tasks, the chi-square scores for the T-CNNs were only marginally higher than those of the P-CNN (Table 2). This was particularly notable for the site identification task, as the T-CNN embeddings were derived from networks explicitly trained for site classification, underscoring the strength of the P-CNN.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Chi-squared scores (x 10³) generated from contingency tables of each recording’s true class and the cluster that each recording was assigned to by affinity propagation clustering. Higher scores indicate that clusters better represented true classes. All chi-square scores were highly significant (p < 0.001).

https://doi.org/10.1371/journal.pcbi.1013029.t002

3.2 Predicting habitat class and site identity with supervised learning

Overall, the classifiers performed well in predicting habitat and individual site classes from one-minute reef soundscape recordings. The six tasks exhibited a range of difficulty, with a mean accuracy across all tasks ranging from 0.56 (±0.05) to 1.00 (±0.0) corresponding to a performance above random classification ranging from 0.23 (±0.11) to 0.87 (±0.0) (Table 3 and S7 Fig). The strongest performance at site classification tasks was reported for the French Polynesian dataset and the weakest was reported for the Australian dataset. However, we find the training and test divisions selected for a given repeat strongly influenced results. For example, the French Polynesian habitat classification task had a mean accuracy of 0.83 using the T-CNN across repeats, but the standard deviation was high (±0.2), with some repeats failing to improve beyond random classification.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Mean and standard deviation of classifier accuracy across repeated training instances using each the three machine learning methods (compound index, pretrained CNN and trained CNN) at six different tasks. Accuracy is the proportion of one-minute recordings from the test data that were correctly classified. Methods where accuracy was reported as significantly higher by the ANOVA test are indicated in superscript next to the mean value for the respective method (A = highest group, B = second highest group, no letter = lowest group). The Random baseline accuracy indicates the expected accuracy of a model that performs random classification. N = 100 for all tasks, except the Fish diversity (Australia) and Depth (French Polynesia) tasks, where N = 32.

https://doi.org/10.1371/journal.pcbi.1013029.t003

The three embedding approaches reported comparatively similar accuracies to one another (Table 3). ANOVA tests reported no significant difference between any of the three methods for the habitat identification tasks set from the Australian and French Polynesian datasets (Table 3). However, significant accuracy differences were observed between methods for the Indonesian habitat classifier and for site identification classifiers across datasets. For three of the four tasks where a difference was observed, the T-CNNs outperformed the other two methods, including the habitat and site classifiers for the Indonesian dataset and the site identity classifier for the French Polynesian dataset. However, the compound index had significantly greater accuracy than the P-CNN and T-CNN for the Australian site classifier.

Inspection of the confusion matrices used to interpret classifier performance revealed that the three embedding extraction methods reported similar patterns of misclassification for each task, where samples from any given class were assigned to the same incorrect classes across methods (S8A, S8B and S8C Fig). The only exception was the French Polynesian habitat classification task where the random forests trained on the compound index and P-CNN were more likely miss-classify shallow samples as mesophotic, whereas the T-CNN was more likely to misclassify mesophotic samples. The only instance where a class was misclassified the majority of the time across all repeats was for site A in the Australian dataset, where samples were frequently assigned to class G (S8B Fig). This misclassification was likely due to the close proximity of these sites (S1 Fig) and their habitat attributes, both of which were ‘low’ fish diversity sites.

3.3 Benchmarking against current automation: acoustic indices

Statistical comparisons of individual acoustic indices represent the predominant method for automated analysis of reef soundscape data [19,44,45]. To assess the performance of individual indices, we deliberately selected those that showed the most significant differences between habitat classes. These were the full-band acoustic complexity index (ACI), the normalised difference soundscape index (NDSI), and the low-band acoustic complexity index (ACI), respectively (Fig 3), which reported Mann-Whitney U test results of: U = 2.00 x 10⁶, p < 0.0001, U = 7.71 x 10⁶, p < 0.0001, and, U = 3.64 x 10⁶, p < 0.0001. Despite this biased selection, our analysis showed that these indices were still unable to classify individual recordings with the same accuracy as ML methods (Fig 3). The proportions of recordings that could be unambiguously classified as the correct habitat type were only 5.2%, 13.0%, and 12.2% for each dataset respectively. Even the lowest-performing ML classifiers, across all tasks and repeats, achieved accuracies of 58.1%, 57.7%, and 47.1% for each embedding type (Table 3). Furthermore, individual acoustic indices failed to classify the site of origin for the recordings (S9 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Boxplots of the acoustic index for which the highest significant difference between habitat classes was reported for the (A) Indonesian, (B) Australian and (C) French Polynesian datasets.

Boxes and their bars represent the 25th, 50th and 75th quartiles.

https://doi.org/10.1371/journal.pcbi.1013029.g003

4.1. Machine learning unlocks greater insights from coral reef soundscapes

Using three biogeographically independent datasets, we demonstrate the potential of machine learning (ML) to automate the large-scale analysis of coral reef soundscape data (20,437 one-minute recordings in the present study). Previous work using automated techniques has predominantly used individual acoustic indices, which can reveal relationships between the soundscape and ecological attributes on reefs [10,33,46–48]. However, comparing individual indices provides only a weak signal [2,28,49,50], which is supported by our findings (Figs 3 and S9). Instead, we show that combining features into embeddings and using these as inputs for ML algorithms provides a powerful alternative, especially when using embeddings output by deep-learning models. We find unsupervised learning algorithms can be used to effectively explore and understand trends within these data. However, to ensure the reliability of ML-based soundscape analysis, careful consideration must be given to sampling design and data partitioning to mitigate biases from temporal variability, instrumentation, and other influences [51].

Looking to the ecological properties of our datasets, these results further support the hypothesis that the raw soundscape of coral reefs shares a relationship with properties such as fish community composition and coral cover (Fig 2 and Table 3). This relationship likely reflects variations in the soniferous community, primarily fish and sound-producing invertebrates, which are known to correlate with coral cover [8,10,33], despite corals themselves not contributing directly to the soundscape. We also provide further evidence that shallow and mesophotic reefs often exhibit distinct soundscapes [45,52–55] (Fig 2 and Table 3). More research is needed to understand the factors that drive this difference, such as the ecological community or oceanographic conditions [56].

Our findings provide further evidence that temporal patterns influence reef soundscapes [11,57], with clusters typically grouping recordings from similar time periods (Figs 2C and S4-S6). For example, crepuscular periods linked separate night and day clusters for the French Polynesian dataset (Fig 2C). Furthermore, ML was often able to differentiate individual sites, an insight which could not have been elucidated using acoustic indices (Fig 3). However, our site identification findings underscore the need to mitigate recorder bias, as models may instead learn to identify properties of recordings from specific devices, a common issue in machine learning analyses. Notably, the French Polynesian dataset exhibited the most pronounced site differences, likely as it was the only dataset in which the instruments could not be rotated (S1 Text).

When comparing the performance across datasets, the unique biogeography, instrument rotation strategy, and sampling regime across time likely interact, making it difficult to disentangle which of these factors had the strongest impact on performance. Unsupervised learning further confirmed that each dataset exhibited unique properties (Fig 2A), likely due to both the soundscape properties and the artefacts of individual recording devices used, highlighting the need for caution when attempting to infer ecological variables using models or findings from another dataset [37].

4.2 Performance and comparison of machine learning approaches

Regarding the selection of the optimal feature extraction approach for ML, a clear difference in performance between the methods was revealed by the unsupervised learning tasks. Here, the T-CNNs produced clusters with a notably greater fidelity to true classes for every task compared to the compound index and in most tasks when compared to the P-CNN (Fig 2C and Table 2), which in turn outperformed the compound index. These results reveal that feature embeddings obtained from deep learning yield improved outputs compared to handcrafted compound index embeddings. Notably, this represents the first demonstration that transfer learning using a network pretrained on data from an entirely different audio domain, with the 5.2m hrs of YouTube audio used to train VGGish, can be successfully applied to coral reef soundscapes. Further improvements to feature embeddings may be achieved by testing alternative DL embeddings developed on large bioacoustic datasets and by incorporating temporal information or architectures that capture long-range dependencies, such as transformers [58,59].

When considering the computational costs, generating feature embeddings with the T-CNN was significantly more computationally intensive than the other two methods. The Indonesian dataset represented the smallest of the three, with 3,335 one-minute recordings. Training each repeat of the T-CNN site classifier on the Indonesian dataset took approximately one-hour to train on an NVIDIA A100, an often prohibitively expensive piece of equipment to access. Training with a CPU available on a standard personal computing device was unable to complete even one of the 50 training epochs in a 24 hour period. Conversely, the computational resources required to run the compound index and P-CNN are much more accessible, taking 84 and 96 minutes respectively to extract embeddings from all recordings in the Indonesian dataset. Following embedding extraction, the execution time of downstream processing using random forest classifiers and unsupervised learning was negligible on a CPU.

When selecting an unsupervised or supervised approach, we find that unsupervised learning offers several advantages over supervised learning. Firstly, the natural overlap that can occur between soundscapes across time is more gracefully reported by unsupervised learning which considers the full dataset through cluster overlap or visualisations (Table 2 and Fig 2), whereas this often leads to a classification error if using supervised learning (Table 3). Additionally, various measures can be derived from unsupervised outputs, such as cluster diversity across recording samples to identify more variable soundscapes. Additionally, percentage overlap in cluster allocations and centroid distances within the embedding space can be used to assess soundscape similarity between sites [29,60]. Prior knowledge on the ecological attributes of certain sites could also be used in a semi-supervised way to infer these from other sites [61].

For supervised classification tasks, our findings show that the difference in performance between the three embedding extraction methods on reef soundscape data was minimal. Multiple algorithms and their respective hyperparameters are available. However, it is well documented that as dataset size grows the choice of algorithm matters less [62]. Our findings support this in the context of the large soundscape datasets used here, given the broadly similar accuracy scores across embedding types (Table 3). Furthermore, raw soundscape datasets of this scale are easily gathered, meaning more computationally intensive embedding types have diminishing returns.

Moving forward, for guaranteed optimal performance, we recommend developing T-CNNs specific to each dataset for use in unsupervised approaches, training them on known attributes of the data (e.g., site classes, as used here). However, the P-CNN offers near comparable performance for orders of magnitude less computational cost. The P-CNN also outperforms the compound index and represents a more standardised tool than this index, which requires curation from an extensive set of indices and parameters that currently vary from study to study [44,49]. The standardised features of a P-CNN facilitate easier comparisons across datasets, whereas a T-CNN would require retraining. We therefore recommend a P-CNN for most use cases. Outputs from clustering can then be used in downstream statistical models, for example testing the relationship between ecological metrics and soundscape diversity of sites which can be indicated by the number of clusters each site is assigned to, or, comparing the overlap between clusters of sites where an intervention has been implemented (e.g., reef restoration) with healthy and degraded baselines (see the tutorial in the Code Reporting section).

4.3 Future directions

There are multiple ways future research could build on the findings from this study. Firstly, soundscape data alone can be used for rapid comparisons between different treatments in novel contexts (e.g., restoration, degradation) without requiring additional biodiversity metrics [20,59]. However, integrating standardised ecological survey data across diverse settings and biogeographies is essential for identifying the underlying drivers. Key areas for investigation include determining how reliably the soundscape can serve as an indicator of soniferous taxa, their relationship with broader ecosystem functioning, and how this changes over time [45,48,57]. Such data would also help address the black-box challenge by linking soundscape properties to ecological attributes. While the tasks examined in this study were achievable with our methods, these relationships and their limitations could be better explored by integrating multiple ecological measures and assessing them across gradients rather than through discrete binary classification challenges we devised. For example, future studies could use metrics such as fish diversity or coral cover from sites spanning biogeographical or human pressure gradients. Metrics with the strongest relationships to the soundscape could be identified and used to predict these on new sites using low-effort PAM data collection. Crucially, prospective users should follow best practices when assembling datasets, ensuring control over confounding variables that may affect machine learning algorithms, such as instrument bias, temporal autocorrelation, and careful curation of training and test sets [63–65].

The availability of low-cost recording technology [34,66] and rapid insightful analysis now possible using ML have the potential to greatly increase the scale of ecological assessment on coral reefs. With careful sampling design and evaluation protocols in place, artificial intelligence can serve as a powerful tool for processing this data and unlocking new insights. The potential demonstrated here also has implications for other marine and terrestrial habitats where these techniques could be applied. While there remains much to discover in this field, soundscape ecology has the potential to support our understanding, protection and restoration of coral reefs and other habitats around the world.