Mega-Efficient Wildlife Classifier (MEWC) Case Study

Barry Brook (0000-0002-2491-1517)
Jessie Buettel (0000-0001-6737-7468)

Biological Sciences

wildlife camera trap Tasmania detection classification inference field site ecological community mammal bird deep learning artificial intelligence

We provide training, test and inference data for the Mega-Efficient Wildlife Classifier (MEWC) workflow, using an example camera-trap dataset collected and curated by Barry Brook and Jessie Buettel, from Tasmania, Australia. These images are drawn from a variety of environmental contexts (dry and wet temperate eucalypt forest, woodland, and grasslands) using white-, infra-red (IR) and no-glow flash types from Cuddeback, Reconyx, Swift and Bushnell cameras. The following species or aggregated classes are represented in the dataset: Tasmanian Pademelon (Thylogale billardierii), Bennetts Wallaby (Notamacropus rufogriseus), Tasmanian Devil (Sarcophilus harrisii), Feral Cat (Felis catus), Bare-nosed Wombat (Vombatus ursinus), Brushtail Possum (Trichosurus vulpecula), Fallow Deer (Dama dama), Southern Brown Bandicoot (Isoodon obesulus), Currawong (Black: Strepera fuliginosa, Grey: S. versicolor) and Bronzewing (Brush: Phaps elegans, Common: P. chalcoptera). The latter two classes are birds, each of which consist of an aggregation of two species within a genus; the former eight are mammals. For implementing the classifier training, we provide 4,000 train and 1,000 test images for each of 10 different classes, for a total of 50,000 expert-labelled snips (each sized at 600- × 600-pixel, after being extracted from their original images using the MEWC-Snip tool). For demonstrating the detection, inference, and post-processing pipelines (EXIF writing and image sorting), we provide a sequence of 100 images for each of four field cameras that were not used in training, located on the lead author’s rural property in southern Tasmania: C3: IR flash, C7: white flash, C15: no-glow flash and C21: inbuilt IR flash. Other than the target wildlife, these images include some representations of blank images, humans (the lead author), and vehicles (trail-bike motorcycle), to demonstrate the four broad classes designated by the MegaDetector prior to classification on the animal images.


Camera trapping (field cameras with passive IR detection sensors and white or IR flash units) from sites across Tasmania, Australia.

410407 Wildlife and habitat management 461103 Deep learning
180601 Assessment and management of terrestrial ecosystems 180606 Terrestrial biodiversity

Data Access

  1. training_test_snip_data.tar ({{3055437824 | bytes}})
  2. unprocessed_service.tar ({{692322816 | bytes}})

Brook, B, Buettel, J (2023) Data from: Mega-Efficient Wildlife Classifier (MEWC) Case Study.
10.25959/wm5g-b990 (DataCite reference)
Barry Brook (0000-0002-2491-1517)
James Morse (0009-0003-6198-0460)
Zach Aandahl (0000-0002-9412-8288)