Skip to content
Ecology

Potato for ecology and wildlife monitoring

Review MegaDetector camera-trap detections and BirdNET or Raven call boxes, then export Camtrap DP, Darwin Core, or Raven files from your own server.

Quick startOpen the Study Designer

Potato is the human review step after a camera-trap detector or a bird-call classifier, where people confirm and correct MegaDetector, BirdNET, Raven, or Audacity detections before export as Camtrap DP, Darwin Core, MegaDetector JSON, or Raven and Audacity files. Potato is free and open source and runs on your own computer or server, including a field-station laptop with no internet connection.

Detectors and classifiers have made camera-trap and acoustic surveys feasible at scale, and most projects still check the output by hand. An evaluation of three AI platforms for camera-trap data found that their species classifications generally had low to moderate recall (Vélez et al., 2022), so a person usually verifies what the model reports before the output becomes data. Potato runs that verification. Potato reads the detector's output and does not run the detector, so you run MegaDetector, a species classifier, or BirdNET-Analyzer on your own machine first and import the files the detector writes.

Ecology tasks in Potato

Each row below is a review workflow that ships with Potato 2.12, and the linked page gives the commands and options.

TaskHow Potato does itDocs
Triage camera-trap imagesA radio question pre-filled with the detector's top category (empty, animal, person, vehicle), answered with the keys 1 to 4Camera traps
Correct detection boxes and speciesBoxes pre-drawn from the detector at or above a confidence threshold, and a species multiselect when the source names speciesCamera traps
Review whole sequencesImages that share a seq_id shown as one item, as a gallery of framesCamera traps
Verify BirdNET detectionsEach recording as one item, with BirdNET's boxes drawn on the spectrogram and dashed until someone changes themBioacoustics
Draw call boxes from scratchA spectrogram_annotation scheme with time-frequency boxes and labelsSpectrogram annotation
Map sightings or habitatsA geo_annotation scheme with points, polygons or region choices on your own GeoJSON mapMaps

Reviewing MegaDetector output

The import builds a project from a MegaDetector batch file and the images MegaDetector ran on. With --image-dir, Potato serves the images from their existing folder, so a large survey does not have to be copied.

bash
potato import -i md_output.json --image-dir /data/traps --md-threshold 0.5 -o review-project/
python potato/flask_server.py start review-project/config.yaml -p 8000
python -m potato.export -c review-project/config.yaml -f camtrap_dp -o package/

--md-threshold drops detections below that confidence before anyone sees the detections (the default is 0.2). A higher threshold means fewer boxes to delete and more missed animals to draw, so set the threshold from a sample of your own images. A pre-filled answer is stored only when the annotator saves the image, so an export contains reviewed images only. Boxes keep the detector's name and confidence and are marked as edited when someone changes a box, so you can report how many detections the review changed.

Getting the data out

The exports are written for the tools and repositories ecologists already use, and each one reads the reviewed answers.

  • megadetector writes batch-output JSON with the corrected boxes, for tools that read MegaDetector results.
  • camtrap_dp writes a Camtrap DP 1.0 package (deployments, media, observations and datapackage.json). Coordinates and package metadata come from item fields or --option values, and anything missing is named in the warnings, so run a Camtrap DP validator before publishing.
  • darwin_core writes a Darwin Core occurrence table, with basisOfRecord set to MachineObservation where the model's identification was left unchanged and HumanObservation otherwise.
  • raven and audacity_labels write one selection table or label track per recording, which open next to the audio in Raven or Audacity.

Working alongside Timelapse, Wildlife Insights, and Raven

Timelapse is a desktop program for tagging camera-trap images that reads MegaDetector output. Wildlife Insights is a hosted platform that classifies uploaded images and lets users review the results. Raven Pro is the Cornell Lab's program for viewing and measuring sounds. Potato fits when several people share the review, because Potato assigns images or recordings to reviewers and keeps each person's answers. Potato also fits when the images cannot leave your own server, because the whole tool runs on that server with no outside connection once installed.

Frequently asked questions

Can I load MegaDetector output and review only the images it flagged?

Yes. potato import reads a MegaDetector batch file, and --md-threshold drops detections below a confidence you choose, so an image with no detection above the threshold starts as empty. Annotators confirm or change that triage answer with one key, then correct the boxes on the images that need correcting.

Can I verify BirdNET detections one clip at a time?

Yes. potato import -f birdnet reads both of BirdNET-Analyzer's output formats, and --min-confidence sets which detections come in. Each recording opens with BirdNET's boxes drawn on the spectrogram, and annotators move, resize, relabel, or delete the boxes and draw the calls BirdNET missed. Potato does not include BirdNET's models, which are licensed CC BY-NC-SA.

Can the output be published to GBIF?

Potato writes the two formats GBIF publishing starts from, a Camtrap DP 1.0 package with camtrap_dp and a Darwin Core occurrence table with darwin_core. Publishing itself happens outside Potato, and the Camtrap DP export lists any metadata the exporter could not fill in.

Can Potato run on a field-station laptop with no internet?

Yes, once Potato is installed. Potato serves its whole interface from the install and needs no connection to review images or recordings that are on the machine. Air-gapped deployment lists what is served locally.

Further reading

References

  1. Vélez et al. (2022). An evaluation of platforms for processing camera-trap data using artificial intelligence. Methods in Ecology and Evolution. https://researchonline.ljmu.ac.uk/id/eprint/19506/

Start here

Camera trapsReview MegaDetector output and export Camtrap DP or Darwin Core.BioacousticsCorrect Raven, Audacity, or BirdNET boxes on a spectrogram.Spectrogram annotationThe time-frequency box scheme and its options.MapsPoints, polygons, and regions on a map, exported as GeoJSON.Detecting rubber-stamped pre-labelsCheck whether reviewers really looked at the detector's boxes.Aggregating crowd labelsCombine several volunteers' answers into one label.

Example designs

Each design is a working Potato configuration you can download and run.

iWildCam Wildlife Detection & ClassificationCamera trap image classification for wildlife monitoring (Beery et al., CVPR 2019). Classify wildlife species from camera trap images across diverse ecosystems worldwide.CUB-200-2011: Fine-Grained Bird Species DatasetCUB-200-2011 is a fine-grained bird image dataset of 11,788 photos across 200 species, each labeled with parts, attributes, and a bounding box (Wah et al., 2011). This Potato config reproduces the labeling task.BigEarthNet Remote Sensing ClassificationMulti-label land cover classification from Sentinel-2 imagery (Sumbul et al., IGARSS 2019). Classify satellite patches into 43 land cover classes following the CORINE taxonomy.FLAIR: French Land Cover from Aerospace ImageryLand use and land cover classification from high-resolution aerial imagery. Annotators classify the primary land use category of aerial image patches and identify any secondary land uses present. Based on the FLAIR dataset from the French National Institute of Geographic and Forest Information (IGN).

Other fields

All use cases