Fig 1.
Percent of snake species represented in the training dataset for each country.
The map was drawn using R package rworldmap v 1.3.6 [26]. See S1 Data for underlying values.
Table 1.
Performance comparison for two ViT-Base/32 models trained on the SnakeCLEF2021 dataset and its “clean” subset.
Fig 2.
The first image from the left is the original image. Top: image by chiuluan, iNaturalist 207060926 (CC-BY); Bottom: image by Alex Karasoulos, iNaturalist 207359674 (CC-BY).
Table 2.
Test time augmentations experiment.
Table 3.
Achieved performance on the SnakeCLEF2021 test set using different locational data and metadata integration method.
Table 4.
Overall performance of the model.
Fig 3.
Distribution of F1 scores according to the number of training images per species.
See S2 Data for underlying values.
Fig 4.
Maps of average F1 score by country (A All species, B MIVS only).
Countries and territories with the lowest overall F1 scores for all species are Martinique (36%), Haiti (40%), Dominica (62.5%), St. Lucia (62.5%), Aruba (68%), and Papua New Guinea (74.8%). The map was drawn using R package rworldmap v 1.3.6 [26]. See S3 and S4 Data for panels A and B underlying values.
Fig 5.
Scatterplot of Asian and African countries by coverage and F1 score.
See S5 Data for underlying values. For other continents, see S3 Fig. and interactive online version at https://chart-studio.plotly.com/~amdurso/6/#/.
Fig 6.
Comparison of F1 score distribution between non-MIVS (574 species) and MIVS (198 species).
See S6 Data for underlying values.
Fig 7.
Computer vision model performance for identifying lookalike venomous and non-venomous snake species in Southeast Asia (Bungarus spp. vs. Lycodon spp).
All species in these genera from the training data were included. A Confusion matrix of the classification results of the model on lookalike snake species with the percentage (blue colour code) of correctly (diagonal cells) and incorrectly (off-diagonal cells) classified images, B Geographic range of the relevant snake genera based on [25], C Representative photos of some of the lookalike snake species tested with the model. The map was drawn using R package rworldmap v 1.3.6 [26]. See S7 Data for panel A underlying values.
Fig 8.
Computer vision model performance for identifying lookalike venomous and non-venomous snake species in Sub-Saharan Africa (Bitis spp. vs. Echis spp. vs. Causus spp. vs. Dasypeltis spp.).
All species in these genera from the training data were included. A Confusion matrix of the classification results of the model on lookalike snake species with the percentage (blue colour code) of correctly (diagonal cells) and incorrectly (off-diagonal cells) classified images, B Geographic range of the relevant snake genera based on [25], C Representative photos of some of the lookalike snake species tested with the model. The map was drawn using R package rworldmap v 1.3.6 [26]. See S8 Data for panel A underlying values.