Fig 1.
A digital image of an herbarium specimen.
The plant and original label containing the original metadata information, as well as additional reference information for the digital image: a ruler for scale and a color reference grid.
Fig 2.
Overview of the confidence-based workflow.
By only considering labels over a certain probability threshold, we increase the final accuracy of the model at the cost of coverage on the overall dataset (red: wrong label, green: true label, gray: rejected label).
Fig 3.
Rejection/Accuracy curves for all four models.
Associated labels: (a) Budding, (b) Flowering, (c) Fruiting, (d) Reproductive.
Fig 4.
DoY estimation error and number of empty species estimates as a function of the confidence threshold.
Fig 5.
Comparison of study ground-truth and AI estimate, per species, with a confidence threshold of 0.5.
Blue: study ground truth with mean/std, red: AI estimate with mean/std.
Fig 6.
Comparison of study ground-truth and AI estimate, per species, with a confidence threshold of 0.99.
Blue: study ground truth with mean/std, red: AI estimate with mean/std.
Fig 7.
Rejection/Accuracy curves of the INaturalist2018 multi-class classifier.
Table 1.
Expended accuracy results on validation set of INaturalist2018 with confidence thresholds.
Accuracy and rejection rate percentages for particular minimum confidence thresholds.
Fig 8.
Workflow to automatically obtain a macrophenological analysis of flowering time shifts: from non-annotated samples to regional trends.
Fig 9.
Flowering shifts distribution per growth form status.
Fig 10.
Flowering shifts distribution per native/introduced status.
Fig 11.
Flowering shifts distribution per wetland status.
Fig 12.
Flowering shifts distribution per early/late season status.
Fig 13.
Flowering shifts distribution per season spread status.
Table 2.
Results of Welch’s T-test evaluating effect of plant characteristics on flowering time shift.
Table 3.
Results of Pagel’s test for correlated evolution between two binary characters.
Fig 14.
Phylogenetic tree with seasonality character and detected flowering shifts.
Table 4.
Accuracy results on validation set.
Table 5.
Extended accuracy results on validation set with confidence thresholds.
Accuracy and rejection rate percentages for particular minimum confidence thresholds.
Fig 15.
2D projection of the Flowering classifier’s embeddings for the validation dataset from a t-SNE algorithm.
(a) Blue and Green are the Flowering/Not Flowering classes respectively. (b) Additional overlay with the confidence scores.
Fig 16.
Examples of images with Sure and Unsure Flowering classification.
(a,b) High Flowering Confidence (a: 97%, b: 94%), (c,d) High Not-Flowering Confidence (a: 98%, b: 99%), (e, f) Unsure Flowering (e: 41%, f: 62%).
Table 6.
Overview of the classes distribution and amount of data in the INaturalist2018 dataset.
Fig 17.
Model annotation estimated trends for each threshold (Black, Brown, Red, Orange, Yellow, Green, Blue), human annotations estimated trend (Purple).
Fig 18.
Amount of slope error (between the automatically annotated and the human annotated ground-truth) against the number of available samples to perform the regression.
Each dot corresponds to a given species for a given threshold.
Table 7.
National Wetland Plant List Wetland Indicator Status categories and definitions.
See [78] for more in-depth definitions and other information.