A dataset for benchmarking Neotropical anuran calls identification in passive acoustic monitoring

Abstract

Global change is predicted to induce shifts in anuran acoustic behavior, which can be studied through passive acoustic monitoring (PAM). Understanding changes in calling behavior requires automatic identification of anuran species, which is challenging due to the particular characteristics of neotropical soundscapes. In this paper, we introduce a large-scale multi-species dataset of anuran amphibians calls recorded by PAM, that comprises 27 hours of expert annotations for 42 different species from two Brazilian biomes. We provide open access to the dataset, including the raw recordings, experimental setup code, and a benchmark with a baseline model of the fine-grained categorization problem. Additionally, we highlight the challenges of the dataset to encourage machine learning researchers to solve the problem of anuran call identification towards conservation policy. All our experiments and resources have been made available at https://soundclim.github.io/anuraweb/ .

Fig. 1. Overview of the AnuraSet methodological workflow that encompasses the process of dataset creation and benchmarking.

It begins with the collection of passive acoustic monitoring data from four sites in the Neotropics. Subsequently, we annotated the recordings with both weak and strong labels. Leveraging these annotations, we undertook a preprocessing of the data to construct a machine learning-compatible dataset. For solving the problem of anuran call identification, we frame the problem as a multilabel classification challenge, and to establish a baseline model, we adopted a transfer learning approach. Furthermore, we merged a specific task with the dataset, culminating in the creation of a benchmark.

Fig. 2

Data collection of calling activity of Neotropical anuran communities. (a) Geographic location of the four sites where the passive acoustic monitoring data was collected. Sites at the Cerrado biome, INCT17 and INCT41 (dots), and at the Atlantic Forest biome, INCT20955 and INCT4 (squares). (b) Photograph of INCT4 monitoring site at the Atlantic Forest biome. (c) Details of the acoustic sensor used to record anuran calls at the edge of a water body.

Fig. 3

Strong labeling process of raw data using audio and spectrogram. (a) Example of strong labeling. For each 1-min raw audio file sampled, the herpetologist annotator identified and selected the temporal limits of the advertisement call. We annotated calls using separate time selections when spaced more than 1 second apart. (b) Image of an individual of Boana lundii its advertisement call coded as BOALUN in the annotation process (c) A calling male of Boana albopunctata (BOAALB).

Fig. 4. An illustrative example of the advertisement call of Physalaemus albonotatus for the three audio quality categories.

(a) High-quality call (H) shows a high signal-to-noise ratio, no overlap with other sounds, a well-identifiable structure on the spectrogram, and can be easily visualized on the oscillogram. (b) Medium-quality call (M) can be visually identified on the spectrogram but may overlap with other sounds that can be difficult to identify in the oscillogram. (c) Low-quality call (L) shows a low signal-to-noise ratio, is partially masked by other sounds, appears with low intensity on the spectrogram, and cannot be easily identified on the oscillogram.

Fig. 5. Frequency distribution of 3-second samples per anuran species.

The long-tailed distribution is a typical distribution of a real-world species diversity dataset. We split species into the classes of ‘common’, ‘frequent’, and ‘rare’ to determine the effect of sample size on the performance of the species identification problem. Additionally, the occurrence of the same species in different sites is represented by different colored squares at the bottom of the histogram. The training and test set distributions obtained by using the split strategy are depicted with black and blue lines, respectively.

Fig. 6. Performance for benchmarking the species identification problem.

Using the ResNet152 model, we evaluated the species identification problem (see section ‘Experimental Setup’) in each site. The x-axis is the number of samples in the logarithmic scale and the y-axis is the F1-score. Across sites, we found a positive relationship between samples and performance.

Fig. 7. Analytical challenges of the AnuraSet.

(a) spectrogram showing eight different species that were recorded calling in less than eight seconds, highlighting the degree of co-occurrences, call overlap, and fine-grained identification; (b) spectrogram showing a dense chorus with a low signal-to-noise ratio and high sound masking. (c) spectrogram showing the richness and co-occurrence of sounds from different taxa (silhouettes from bottom to top depicting frogs, birds, and orthopterans, respectively).