Acoustic Classification of Insects
Code, trained models and metadata accompanying the paper Optimizing Spectrogram Front-Ends and Training Strategies for Deep Learning-Based Acoustic Insect Classification (Peter & Kunzelmann, submitted). The models classify insect sounds (451 species of Orthoptera and Cicadidae) and are trained on InsectSet459 (Faiß & Stowell, 2025). Contents Code ( src/ , scripts/ , configs/ ): data download and preparation, training and evaluation (PyTorch, EfficientNetV2-S). Trained models ( models/ ): best_model.pth (64 mel bands, mixup, class-weighted loss) and baseline_model.pth (128 mel bands, no mixup, no class weighting). The checkpoints contain the m
Resource information
- Format
- zip
- Original repository
- 23981805
- DOI
- 10.5281/zenodo.23981805
- Credit basis
- Creators in Zenodo record metadata
- Access
- Open files available from the original Zenodo deposit
- Last checked
- Nov 5, 2026
- License
- mit-license
- Last record update
- Nov 5, 2026
About this resource
Code, trained models and metadata accompanying the paper Optimizing Spectrogram Front-Ends and Training Strategies for Deep Learning-Based Acoustic Insect Classification (Peter & Kunzelmann, submitted). The models classify insect sounds (451 species of Orthoptera and Cicadidae) and are trained on InsectSet459 (Faiß & Stowell, 2025). Contents Code ( src/ , scripts/ , configs/ ): data download and preparation, training and evaluation (PyTorch, EfficientNetV2-S). Trained models ( models/ ): best_model.pth (64 mel bands, mixup, class-weighted loss) and baseline_model.pth (128 mel bands, no mixup, no class weighting). The checkpoints contain the model weights without the optimizer state. Metadata ( metadata/ ): every InsectSet459 recording used and not used (with split, licence and contributor), a species table with Catalogue of Life, GBIF and ITIS identifiers, the settings of all 112 training runs of the study, and an EML record. Environments : conda_environment.yml and conda_environment_pinned.yml (exact versions, Linux x86-64). The audio recordings are not included. They can be downloaded from InsectSet459 with the scripts in scripts/ ; see the README for the full workflow. Licences The licence of this record (MIT) applies to the code and configurations. Other parts of the archive have their own licences: Code and configurations: MIT Metadata ( metadata/ ): CC BY 4.0 Model weights ( models/ ): CC BY-NC-SA 4.0 (non-commercial, as 68% of the training recordings carry a non-commercial licence) The recordings remain under the licences of their original sources, listed per file in metadata/files_used.csv . Citation Please cite the paper (reference will be added once published) and this software.