Deep learning of pollen images under low annotation costs: joint optimization of morphological features and training and prediction strategies

Pollen identification is of great importance in the fields of palynology, palaeoecology, botany, medicine and forensic science, but traditional microscopic morphological analysis methods are inefficient and subjective. In this study, we propose an innovative approach based on deep learning to improve the accuracy and efficiency of pollen identification. We constructed a high-quality pollen dataset containing 5521 images of 141 species and a structured attribute table containing 20 standardized morphological features. With an improved ResNet50 architecture, the model utilizes a masking mechanism to combine image features with morphological data, significantly improving classification performance. In addition, we propose a joint training strategy that utilizes both weakly labeled data (unlabeled images + some morphological features) and fully labeled data to alleviate the data scarcity problem. The experimental results show that with the introduction of morphological features, the accuracy of the model significantly improves from 83.00% to at least 89.49% and exhibits stronger generalization ability, effectively reducing overfitting. This study provides a scalable solution for automated pollen identification, addressing key challenges in data utilization and classification accuracy.