AnuraSearch (formerly FrogNET): Acoustic Classifier and Training Dataset for 11 European Anuran Species

This repository contains the dataset, machine learning model files, and reproducibility package associated with the research article: "Repurposing bird-based acoustic models for amphibian surveys: Strengths and limitations in real-world deployments". Concept DOI (always resolves to the latest version): 10.5281/zenodo.17736510 The repository provides: AnuraSearch Model: An automated acoustic classifier built on a frozen BirdNET v2.4 feature extractor (Kahl et al., 2021) with a custom classification head optimized for 11 Eastern European anuran species. Training/Validation Dataset: Curated 3-second audio clips sourced from open-access bioacoustic archives (Xeno-canto and Observation.org) used for model development and validation. Reproducibility Package: Complete Python analysis code, raw detection files, reference metrics, and a step-by-step protocol to reproduce all evaluation metrics, tables, and figures reported in the article. Version 3 Updates (3 September 2026): Updated Reproducibility Package: Contains refined analysis scripts, reference outputs, and verification routines (the model files and training dataset remain unchanged). Refined Non-Detection Evaluation: Threshold-dependent metrics are computed over [0.05, 1.00], representing the classifier's actual confidence range. Clips without detections carry no confidence score and contribute directly as False Negatives or True Negatives. Confidence Calibration Analysis: Includes new evaluation scripts and diagnostic outputs (Tables S3–S4, Figures S-cal-1–S-cal-2) analyzing confidence score calibration, identifying score saturation (>0.75) and assessing detection-only reliability under field conditions (detection-only AUROC = 0.582, 95% CI: 0.554–0.611). Persistent Identification: Internal documentation and citations have been standardized to reference the Concept DOI (10.5281/zenodo.17736510) to ensure version-proof citations. For detailed execution instructions, verification steps, and a complete version history, please refer to README.txt, PROTOCOL.md, and CHANGELOG.md inside the reproducibility package.