Akademik Veri Yönetim Sistemi
The Journal of Apitheraphy and Nature, cilt.8, sa.2, ss.268-294, 2025 (TRDizin)
Melissopalynology is the gold standard for authenticating honey but
traditional microscopic analysis is time-consuming and subjective. This
study evaluates a hybrid artificial intelligence approach to automate
pollen classification using the comprehensive POLLEN73S dataset, which
features 73 distinct pollen types from the Brazilian Savanna. To address
class imbalance, the dataset was expanded to 7300 images using data
augmentation. We extracted morphological features using three
pre-trained deep learning models (ResNet50, EfficientNetB0, MobileNetV2)
and classified them using 17 traditional machine learning algorithms.
The hybrid model combining ResNet50 features with Linear Discriminant
Analysis (LDA) achieved the highest accuracy of 97.00%. Error analysis
indicated that misclassifications were concentrated among taxonomically
similar genera, such as Serjania, due to shared exine structures. These
results demonstrate that the proposed hybrid model offers a highly
accurate and scalable solution for laboratory-based honey
authentication, provided it is integrated with debris detection systems
to handle real-world samples.