frognet

This project was initiated to attempt to replicate the results of BirdNET, but with frogs in Missouri. We've only coded the portion to obtain calls from iNaturalist, split them into 3 second samples and manually review them for use. That code resides in frog_calls_begin.py and is set up to run in individual cells in a Jupyter notebook. Cells at the end are for diagnostic purposes and will go away as the project nears completion.

Setup

Requires Python 3.12+. Install dependencies with:

pip install .

Then launch Jupyter and open frog_calls_begin.py as a notebook (e.g. via the Jupytext extension or by converting it first).

Workflow

Run the cells in order for each species:

1. Setup — configure CURRENT_SPECIES and create the data/ directory structure
2. Load observations — reads data/downloads/<species>/full_calls.csv (exported from iNaturalist using a query for RG observations of the species with sound within Missouri, or surrounding states if Missouri doesn't have enough observation, and renamed to full_calls.csv)
3. Download sounds — fetches audio files from iNaturalist observation URLs
4. Segment audio — splits each recording into 3-second WAV clips
5. Generate spectrograms — saves a PNG spectrogram for each clip
6. Score clips — auto-classifies clips as keep, discard, or review using signal band amplitude or RIBBIT (for trill species)
7. Manual review — plays borderline clips for human keep/discard decisions
8. Export confirmed — copies all keep clips to data/snippets/<species>/confirmed/

Frog Classifier

data/frog_classifier.tflite is a custom BirdNET classifier trained on the confirmed clips listed below. It recognizes 24 Missouri anuran species and is intended to be used as a drop-in replacement for BirdNET's default classifier when analyzing field recordings for frogs.

Companion files:

• data/frog_classifier_Labels.txt — species list in classifier order
• data/frog_classifier_Params.csv — BirdNET training parameters
• data/frog_classifier_sample_counts.csv — per-species confirmed clip counts used during training

To run the classifier on a folder of WAV files:

./run_frognet_model_on_files_in_folder data/field_recordings/<folder>

This calls birdnet-analyze with settings tuned for frog calls (100–7000 Hz, min confidence 0.1, combined CSV output). Results are written to the same folder as BirdNET_CombinedTable.csv and per-file .csv files.

Spectrogram Tool

spectrogram_best_detections.py reads the BirdNET_CombinedTable.csv produced by the classifier and generates one spectrogram PNG per detected species — the clip with the highest confidence score. Each image is labeled with the source filename, species name, and confidence value.

python spectrogram_best_detections.py data/field_recordings/<folder>

PNGs are written to the same folder as the CSV.

Training Data — Confirmed Clips per Species

10,651 confirmed 3-second clips across 24 species, all sourced from iNaturalist Research Grade observations.

Species	Confirmed clips
Acris blanchardi	303
Anaxyrus americanus	552
Anaxyrus cognatus	117
Anaxyrus fowleri	383
Anaxyrus woodhousii	168
Dryophytes chrysoscelis	286
Dryophytes cinereus	203
Dryophytes versicolor	636
Gastrophryne carolinensis	245
Gastrophryne olivacea	841
Lithobates areolatus	210
Lithobates blairi	140
Lithobates catesbeianus	326
Lithobates clamitans	286
Lithobates palustris	148
Lithobates pipiens	193
Lithobates sphenocephalus	583
Lithobates sylvaticus	986
Pseudacris crucifer	848
Pseudacris feriarum	1,419
Pseudacris fouquettei	893
Pseudacris maculata	603
Scaphiopus holbrookii	215
Spea bombifrons	67
Total	10,651