Skip to content

Identifying Species with AI

Biowatch runs species identification models locally — images are analyzed on your machine and never uploaded anywhere.

Installing Models

Open Settings (bottom of the studies sidebar) → AI Models. The coverage map shows which model fits your study region; the cards below it let you download and manage each model.

AI model settings
The AI Models settings page with the coverage map

Click Download on a model card to install it (sizes range from ~120 MB to ~1.2 GB — see the table below). Once downloaded, the model appears in the model dropdown of the Images Directory import, and Delete reclaims the disk space whenever you no longer need it.

Model list
All available models with their coverage, size, and species count

Available Models

Model Provider Coverage Species Download size
SpeciesNet Google Worldwide 2,000+ ~470 MB
MegaDetector Microsoft AI for Earth Worldwide Detection only (animal / person / vehicle) ~120 MB
DeepFaune CNRS Europe 26 ~1.2 GB
Manas OSI-Panthera & Hex Data Himalayas (Kyrgyzstan) 11, incl. snow leopard ~500 MB

Which one to pick?

  • SpeciesNet is the safe default: launched by Google in 2025, it classifies over 2,000 species labels worldwide.
  • MegaDetector doesn't identify species — it finds animals, people, and vehicles. Use it as a fast blank-filter before manual annotation, anywhere in the world.
  • DeepFaune is the best choice for European fauna, developed by CNRS specifically for European camera traps.
  • Manas specializes in Central Asian mountain fauna, with a focus on snow leopards.

Regional models are usually more accurate within their range; the worldwide models are a flexible baseline.

Need a model for your region?

The team can train custom models for specific regions or integrate a model you already have — use Get in touch at the bottom of the AI Models page.

Running a Model

Models run when you import an images directory: pick the model, select the folder, and Biowatch detects and identifies the animals as it builds the study. Predictions land as regular observations — review and correct them in the gallery viewer.