Computer Science & AI · Level 2 · Intermediate
The successes and pitfalls: Deep-learning effectiveness in a Chernobyl field camera trap application
Maile RE, Duggan MT, Mousseau TA (2023). Ecology and Evolution. doi:10.1002/ece3.10454 · CC BY 4.0
Step 1 of 2Read the paper
Can a neural network trained on a fairly small set of photos reliably identify wildlife in camera-trap images, and do rain, darkness, wind or temperature make it fail? The authors trained an image classifier on photos from cameras in the Chernobyl Exclusion Zone, then tested it on 9576 new images and related its successes and failures to weather and light.
How to read it
- Use Table 2 to keep the train, test, validation and case-study image sets straight before reading Results.
- Note how a 'success' is defined in the Case study section; it is stricter than the usual accuracy figure.
- Compare Table 3 (actual vs predicted counts) with Tables 7 and 8 (rain and low light).
- Read the Discussion's limitations paragraph: how many events were analysed, and what cameras were used?
Step 2 of 2Take the test
15 questions · about 40 minutes including reading · pick an answer to see straight away if it’s right. Nothing is saved.
Loading progress…
Paper: Maile RE, Duggan MT, Mousseau TA, Ecology and Evolution (2023), licensed CC BY 4.0. Summary and questions are Research Nest’s own; read the original for the full study.