Comparative Analysis of Methods for Improving the Classification Accuracy of Morphologically Similar Pollen Grain...

Automatic classification of pollen grains in microscopic images supports melissopalynology and honey quality control but is complicated by class imbalance and morphologically similar classes. We evaluated factors affecting classification on the Cretan Pollen Dataset v1 (CPD-1; 20 classes, 4025 images) with YOLOv8s-cls, YOLO11s-cls, YOLO26s-cls, EfficientNet-B3 and a reproduced RCANet. Each configuration was trained with five seeds; hyperparameters and ensembles were selected on validation data only. Augmentation improved the macro F1-score by about 1.2 percentage points; a size-matched unbalanced control showed that this gain came from the larger training set, not from class balancing. The most difficult class, Ceratonia, was misclassified mainly as Vitis and Olea; its F1-score fell from 0.878 to 0.778 when its original training images were reduced from 26 to 10. Focal Loss and Squeeze-and-Excitation blocks did not improve Ceratonia reproducibly (changes in its F1-score from -0.03 to +0.03, not significant after Holm correction), and the YOLO architectures could not be ranked reliably. An ensemble of EfficientNet-B3 and YOLO models selected on validation data outperformed the best single model in all seeds (macro F1-score 0.9720 vs. 0.9631). Several single-run effects were not reproduced with multiple seeds, underlining the need for multi-seed evaluation and validation-only selection.