IMPROVING POLLEN SPECIES CLASSIFICATION USING DEEP LEARNING AND SIZE FEATURES

Abstract

Automated pollen classification is challenging because visually similar species may differ only by subtle morphological characteristics. This thesis investigates deep-learning models for classifying microscopic pollen-grain images from ten plant species using RGB image crops together with two segmentation-derived size measurements: the major and minor axes of the pollen grain. The study builds on work by Jurdell (2025) and uses the same ResNet18-based image-and-size model in Python/PyTorch to establish a comparable baseline.

This thesis evaluates the effects of missing-size handling, validation-based model selection, class-balancing augmentation, image-feature representation, focal loss, and backbone architecture. The reproduced baseline, denoted R512-Base, achieved 76.8% test accuracy and 76.2% macro recall. Species-wise mean imputation of missing size features and the change in image representation improved performance. The best overall model replaced Resnet with ConvNeXt-Tiny, achieved 85.2% test accuracy and 85.3% macro recall.

The improvement was not uniform across species. The most persistent difficulty was the separation of Crepis capillaris and Hypochaeris radicata. In the baseline model, Crepis Capillaris had zero recall; the final ConvNeXt-Tiny model substantially increased the recall of Crepis capillaris to 38.5%. Thus, the main limitation was not poor performance across all species, but concentrated uncertainty around a small number of visually similar and flower-sensitive classes.
A sensitivity analysis using 16 alternative flower-level splits for Crepis capillaris and Hypochaeris radicata showed that the final model remained strong on average, with mean accuracy 84.2% and mean macro recall 85.2%. However, the recalls of these two species varied substantially depending on which flowers were held out for testing, whereas the remaining species were comparatively stable. These findings support the use of flower-level evaluation for pollen classifiers and show that aggregate accuracy should be reported together with class-wise recall and error-structure analysis.

Popular Abstract

Pollen grains can reveal which flowers bees and other pollinating insects have visited. This information is important for understanding how plants and pollinators interact, and for studying the health of ecosystems. However, identifying pollen species by looking through a microscope is slow, difficult, and requires expert knowledge. As a result, there is growing interest in using artificial intelligence to support and speed up pollen analysis.

This thesis investigates how deep learning can be used to classify microscopic images of pollen grains from ten plant species. Each pollen image was combined with simple size information, such as the major and minor axes of the grain. The major and minor axes refer to standard image-analysis shape measurements. After segmentation, each object is approximated by an ellipse, and the major and minor axes are the longest and shortest principal axes of this fitted ellipse, respectively. This is consistent with common image-analysis outputs such as MATLAB’s regionprops. The idea was that while many pollen species look very similar, their size and shape may contain useful clues that help the model distinguish between them.

Several model settings were tested, including different ways of handling missing size measurements, image augmentation, and alternative neural network architectures. The best-performing model used a modern convolutional neural network called ConvNeXt-Tiny. It improved the test accuracy from 76.8% in the baseline model to 85.2%, showing that better image representations and careful training choices can make pollen classification more reliable.

The results also showed that not all species were equally difficult to classify. Most species were identified with high accuracy, but two visually similar species, Crepis capillaris and Hypochaeris radicata, were repeatedly confused with each other. This suggests that the main challenge is not simply classifying pollen in general, but separating species that are genuinely hard to tell apart, even from microscope images.

A further analysis showed that model performance can depend on which individual flowers are used for testing. This highlights the importance of evaluating pollen classifiers on pollen from unseen flowers, rather than only on randomly mixed images. Overall, the study shows that deep learning can be a useful tool for automated pollen identification, but that reliable ecological use requires looking beyond overall accuracy and carefully studying where and why the model makes mistakes.

• author: Truong, Nancy
• supervisor: Johan Lindström
• organization: Mathematical Statistics
• course: MASK11 20261
• year: 2026
• type: M2 - Bachelor Degree
• subject: Mathematics and Statistics
• keywords: Machine Learning, Computer Vision, Deep Learning, Image Classification, Supervised Learning, Unsupervised Learning, t-SNE, Pollen Classification, Pollen, Pollen Analysis, resnet18, convnext, Ecology
• publication/series: Bachelor’s Theses in Mathematical Sciences
• report number: LUNFMS-4090-2026
• ISSN: 1654-6229
• other publication id: 2026:K25
• language: English
• id: 9244882
• alternative location: https://zenodo.org/records/20187141
• date added to LUP: 2026-06-29 13:39:27
• date last changed: 2026-06-29 13:39:27