Fig 1.
Sample images for each of the 23 pollen types from the Brazilian Savannah flora used in the experiments.
(Scale = 10 μm). a) Anacardiaceae: Schinus terebinthifolius; b-c) Arecaceae: b- Arecaceae; c- Syagrus. d-e) Asteraceae: d) Chromolaena laevigata; e) Tridax procumbens. f) Bignoniaceae: Fridericia florida. g) Burseraceae: Protium heptaphyllum. h) Combretaceae: Combretum discolor. i-j) Euphorbiaceae: i) Croton urucurana; j) Mabea fistulifera. k-n) Fabaceae: k) Anadenanthera colubrina; l) Dipteryx alata; m) Mimosa somnians; n) Senegalia plumosa; o) Lamiaceae: Hyptis. p-q) Myrtaceae: p) Eucalyptus; q) Myrcia. r) Poaceae: Urochloa decumbens. s) Rubiaceae: Faramea. t-u) Sapindaceae: t) Matayba guianensis; u) Serjania laruotteana. v) Urticaceae: Cecropia pachystachya. w) Vochysiaceae: Qualea multiflora.
Fig 2.
Example of the image segmentation process.
a) Image captured under the light microscope and containing several pollen grains. b) Resulting segmented image containing just one pollen grain.
Fig 3.
Part of the questionnaire containing a pollen image and the 23 pollen types options from which the beekeepers should select one.
Fig 4.
Five images of Senegalia plumosa pollen used in the support material.
Beekeepers had access to this kind of reference material to answer the question during the application of the questionnaire. The common name of the Senegalia plumosa pollen type is “arranha-gato” in Portuguese.
Fig 5.
Confusion matrix summarizing the human performance on pollen classification.
The rows represent the true pollen types while the columns indicate how the images have been classified by humans. All the correct answers are in the diagonal.
Fig 6.
Boxplots for the CCR performance of the three feature extractors and the humans.
Among the automatic techniques the highest median value was 66%, very near the median of the human performance, 67%.
Fig 7.
Boxplots for the CCR performance of the automatic techniques considering each pollen type.
These boxplots show that automatic classification varies a lot among different pollen types.
Table 1.
Performance of each combination of techniques using three different metrics: CCR, F-Measure and AUC.
For each combination and metric, the mean and standard deviation values are shown. The best performances for each metric are shown in bold. The same capital letters in the superscripts indicate no statistical difference between feature extractors (rows) as the same lower case letters indicate no significant difference between supervised learning techniques (columns).
Fig 8.
Confusion matrix for the best combinations of feature extractor (CST) and classifier (C-SVC).
The rows represent the true pollen types while the columns indicate how the images have been classified by the computer. All the correct answers are in the diagonal.
Fig 9.
Confusion matrix for the combination of BOW and C-SVC.
The rows represent the true pollen types while the columns indicate how the images have been classified by the computer. All the correct answers are in the diagonal.
Fig 10.
Confusion matrix for the combination the CST+BOW feature extractor and C-SVC.
The rows represent the true pollen types while the columns indicate how the images have been classified by the computer. All the correct answers are in the diagonal.
Fig 11.
Examples of the most easy and most difficult pollen types for human classification.
a) One example of the Chromolaena laevigata pollen that received the highest CCR score for human performance; b) One example of the Qualea multiflora pollen, the hardest pollen during human classification c) Example of the Dipteryx alata pollen that has been mostly confused with Qualea multiflora.
Fig 12.
Dipteryx alata image stacking.
a-d) Images from the same pollen grain. e) Sharp image obtained by stacking the images shown in a-d. To solve the problem of blurred images, the stacking of different focuses of the pollen image can be a promising approach.