Acoustic individual identification in a species of field cricket using deep learning

Emmanuel Kabuga, Diptarup Nandi, Stuart A. Burrell, Gciniwe Dlamini, Rohini Balakrishnan, Bubacarr Bah, Ian DurbachbioRxiv 2025 · DOI

TL;DR

This paper explores using deep learning to identify individual field crickets by their unique chirps. They achieved high accuracy in recognizing crickets, even across different nights and with unseen individuals, demonstrating the potential for non-invasive insect monitoring.

What Problem Does This Paper Address?

Accurately identifying individual animals is crucial for ecological studies, from estimating population sizes to understanding social networks and behavior. Traditionally, this involved physical capture and marking, which is labor-intensive and can be stressful for the animals. More recently, researchers have turned to analyzing passive recordings (audio or video). While this has been successful for species with distinct visual or acoustic signatures (like dolphins or birds), applying it to insects with simpler acoustic signals has been challenging. This paper tackles this gap by investigating the feasibility of using deep learning for individual acoustic identification (AIID) in field crickets, a species with relatively simple, repetitive calls.

Methods

The researchers employed two main deep learning approaches:

  • Standard Classification Neural Networks (CNNs): These models were trained to classify cricket chirps into one of 47 known individual categories. They used spectrogram images of chirps as input.
  • Similarity Learning Neural Networks (Triplet Networks): These models were designed for "open population" scenarios where the number of individuals is unknown. They learn to embed audio samples in a way that minimizes the distance between samples from the same individual and maximizes the distance between samples from different individuals. This allows them to determine if two calls likely belong to the same cricket.

They also compared these deep learning models against Random Forests (RFs), a traditional machine learning method, using both pre-extracted acoustic features (like MFCCs and temporal features) and spectrograms.

Key steps included:

  • Acoustic Signal Segmentation: Identifying individual chirps and extracting fixed-length segments (1-second intervals) from the audio recordings.
  • Feature Extraction: Generating Mel spectrograms (visual representations of sound frequency over time) for CNNs and extracting Mel-frequency cepstral coefficients (MFCCs) and temporal features for RFs.
  • Data Partitioning: Carefully creating training, validation, and testing datasets, including scenarios with individuals seen during training and individuals that were entirely new.

Datasets

The study utilized an acoustic dataset of 137 audio recordings from 47 male field crickets (Plebeiogryllus guttiventris). These recordings were collected over three nights in Karnataka, India. The dataset was partitioned into training, validation, and testing sets to evaluate model performance under various conditions, including:

  • Closed Population: Models trained and tested on known individuals.
  • Open Population: Models trained to compare pairs of calls to determine if they belong to the same individual, including scenarios with individuals not seen during training.

Key Insights and Ideas

  • Insects are not too simple for AIID: This study breaks new ground by demonstrating that even insects with relatively simple, repetitive acoustic signals can be individually identified using deep learning. This opens doors for AIID in a vast array of previously unstudied insect groups.
  • Deep learning excels at harder tasks: CNNs consistently outperformed traditional RF models, especially when faced with more challenging scenarios like classifying calls from individuals recorded on nights not used for training.
  • More data, more robustness: Training models on data from multiple nights significantly improved their ability to generalize and recognize individuals across different recording sessions, highlighting the importance of diverse training data.
  • Temperature matters, but normalization is tricky: Cricket calls are influenced by temperature, affecting chirp duration and frequency. While this introduces variability, simple temperature normalization techniques didn't improve model performance, suggesting that more complex, individual-specific adjustments might be needed.
  • Similarity learning is key for unseen individuals: Triplet networks, designed for open-population scenarios, showed remarkable ability to generalize and identify individuals not encountered during training, making them highly valuable for real-world ecological monitoring.

Key Results

  • Closed Population:
    • Best models achieved up to 99.9% accuracy when trained and tested on calls from the same night.
  • Accuracy dropped to 63.2% when tested on calls from different nights.
  • Pooling data from two nights significantly improved cross-night accuracy by 32.4% for CNNs.
  • Open Population (new individuals):
    • Models identified same-night pairs with 97% accuracy and different-night pairs with 87.1% accuracy.
  • When tested on individuals not seen during training, accuracy was 94.1% within a night and 78.8% across nights.
  • CNNs vs. RFs: CNNs consistently outperformed RF models, especially in cross-night and open-population scenarios.
  • Temperature Effects: Higher temperatures were associated with shorter chirp durations and higher carrier frequencies. Misclassification errors increased with temperature differences, but temperature normalization did not improve performance.

Limitations and Future Work

  • Temperature Normalization: While temperature affects cricket calls, the simple normalization methods used did not improve performance. Future work could explore more sophisticated, individual-specific temperature adjustments.
  • Semi-Automated Approach: The models, while accurate, did not achieve perfect identification for fully automated use. The authors suggest integrating these classifiers with manual processes to create a semi-automated system that can accelerate annotation and maximize efficiency.
  • Broader Insect Application: Given the structural similarity of acoustic signals across many insect species, the developed classifiers could be extended to investigate AIID in other orthopterans, katydids, grasshoppers, and even cicadas.
  • Data Availability: The code and a subset of audio files are publicly available, facilitating further research and development in this area.

Summary updated

Abstract

Abstract Individual animal identification is essential for wildlife conservation and management, helping to estimate animal abundance and related parameters. We assessed the feasibility of identifying individual field crickets ( Plebeiogryllus guttiventris ) from their calls using deep learning. In a closed population of known individuals, the best models recognized crickets with up to 99.9% accuracy when trained and tested on calls from the same night, and 63.2% when tested on calls from nights not used in training. When matching pairs of calls, without knowledge of all individuals in the population, models identified pairs of calls from the same night with 97% accuracy, falling to 87.1% if calls were from different nights. Accuracy remained high when tested on individuals not observed during training (within night: 94.1%; across nights 78.8%). Pooling training data across multiple nights improved test accuracy for all models. Deep learning outperformed random forests, particularly on harder tasks, although both were able to discriminate individuals. Analysis of temperature effects on cricket calling behavior showed that higher temperatures were associated with shorter chirp durations and higher carrier frequencies. However, adjusting spectrograms for chirp duration alone, or for both duration and frequency, did not improve performance. The results are the first to demonstrate AIID using wild recordings of an insect species, and further highlight the potential of deep learning-based AIID for non-invasive monitoring of animal populations.
BibTeX
@misc{kabuga2025acoustic,
  author = {Emmanuel Kabuga and Diptarup Nandi and Stuart A. Burrell and Gciniwe Dlamini and Rohini Balakrishnan and Bubacarr Bah and Ian Durbach},
  title = {{Acoustic individual identification in a species of field cricket using deep learning}},
  year = {2025},
  doi = {10.1101/2025.05.24.655958},
  howpublished = {bioRxiv},
}
Figure