Table 1.
Dataset glossary: This reference table indicates where the data comes from and what it is used for.
Each dataset falls under either the CNN dataset (Non-fossil images used to train, test, and validate the model) or the fossil image dataset (Fossil object images captured using the Classifynder automated slide-scanner; used to further test the CNNs).
Fig 1.
Lac Bélanger, whose organic sediment core is used to generate the full fossil dataset. It is located within Québec’s mixed-temperate forest, which is comprised of both temperate forest species (Pinus strobus, Acer saccharum, Quercus, Betula alleghaniensis) and stress-resistant boreal species (P. banksiana, Abies balsamea, Picea, B. papyrifera).
Fig 2.
It shows the composition of the training data (blue boxes that are grouped into the dark-gray boxes), how the seven different CNNs are hierarchically joined in a full network (yellow boxes and arrows) and their outputs (green boxes). In order to be classified, an unknown image travels top-down through the first CNN (Level 1; Main CNN) and then either be classified into an output class (green boxes) or go through a lower-level CNN (level 2) where the same process would repeat. An image could potentially travel down to the third level (Either Alnus CNN or Acer CNN), thus achieving a finer taxonomical identification.
Table 2.
Information on the training dataset.
Each pollen taxa (Species) makes part of a Class, usually made up of pollen of a similar morphological group. For each taxa, the source, their additional laboratory treatment (if any) and their number of grains is tallied. The training dataset is fed into the model in order to train it (Fig 2).
Fig 3.
Different pollen taxa as captured by the Classyfinder slide scanner.
Note the change in scale between images 1–2, 3–4, and the rest.
Table 3.
This table synthesises the parameters used in the training of the seven CNNs. The “Images per class” parameter accounts for all of the corresponding training, testing and validation images. The “Under threshold” metric corresponds to the percentage of test images that fell under the 0.7 confidence threshold during evaluation i.e. the proportion of the test set that was classified as “unknown” (μ: 4%). The NLL Improvement metric (Negative log likelihood) quantifies the impact of Temperature-scaling using the validation data.
Fig 4.
Architecture of the convolutional neural networks.
The first three layers present after the inputs (left to right) are transfer learning layers. The next six layers represent the common core, present in all seven networks. The following two layers were present only in the deep architecture networks. For the Alnus CNN, a shallower model was used wherein the last convolutional and max pooling layers are removed. The fully connected layer is comprised of the last four layers. See [19] for a review on the terms and layers discussed in this paper.
Fig 5.
The confusion matrix showing the training accuracy of our seven models when tested on the test dataset.
The models are arranged in hierarchical order from the top-left towards the bottom-right. The data has been normalized.
Fig 6.
Proportion over time of each possible output class, minus Alnus crispa and rugosa.
Both the observed n and CNN predicted n are next to their class plot. The time series corresponds to 30 carbon-dated samples analysed along lac Bélanger sediment core.
Fig 7.
Scatter plot of each possible output class, minus Alnus crispa and rugosa.
P-values under 0.05 indicate a correlation according to Student’s t-test. The R2 value indicates the goodness of fit of the regression curve, while the root-mean-square deviation (RMSE) indicates the spread of the data, i.e. if the goodness of fit is affected by outliers.
Table 4.
The results of the statistical tests are listed here.
The class column indent shows the classification level and respective parent classes (if any) of the 18 output classes. n refers to the number of grains identified by the palynologist, while n(CNN) refers to the number of grains as identified by the model. The average per-class accuracy (APC) has only been computed for non-final output classes. 6 out of 18 classes failed Student’s t-test.
Fig 8.
Pollen abundance diagram plotted using the model’s predictions on the full fossil dataset, totalling 271 samples.
When present, the color fills represent a 5x data exaggeration. The total observed amount of Eucalyptus grains was too low to present the data as pollen influx (pollen/cm2/year), so abundance (%) is used instead. The percentage of NPP/minerals per sample is shown as well as the results of the CONISS cluster analysis and its pollen biozones. The dotted lines cutting across the biozones represent the palynologist biozones (Fig 9). The lefthand ticks represent the 14-C dates. See S1 Fig for the age-depth model.
Fig 9.
Pollen abundance diagram plotted using data gathered by a palynologist working on a traditional light microscope.
When present on the diagrams, the color fills represent a 5x data exaggeration. The results of the CONISS cluster analysis and its resulting biozones are also shown. Considering the low A. rubrum abundance, its influx is instead represented as dots.