Pollen Grain Classification Using Some Convolutional Neural Network Architectures

Area of Science:

• Computer Science
• Botany
• Machine Learning

Background:

• Pollen grain classification is crucial for botany and allergy research.
• Existing methods often struggle with accuracy and efficiency.
• Convolutional Neural Networks (CNNs) show promise for image-based classification tasks.

Purpose of the Study:

• To improve pollen grain classification accuracy using advanced CNN architectures.
• To benchmark the performance of eight popular CNN models on a Brazilian savanna pollen dataset.
• To identify the most effective CNN architecture for this specific classification task.

Main Methods:

• Utilized a public dataset (POLLEN73S) of 2523 annotated pollen images from the Brazilian savanna.
• Applied and evaluated eight well-known CNN architectures: InceptionV3, VGG16, VGG19, ResNet50, NASNet, Xception, DenseNet201, and InceptionResNetV2.
• Employed holdout cross-validation for performance assessment.

Main Results:

• DenseNet201 achieved the highest accuracy at 97.217%, surpassing previous results by 1.517%.
• ResNet50 demonstrated strong performance with 94.257% accuracy, an improvement of 0.257%.
• VGG19 yielded the lowest performance among the tested architectures, with 89.463% accuracy.

Conclusions:

• DenseNet201 and ResNet50 are highly effective CNN architectures for pollen grain classification.
• The study demonstrates significant improvements in accuracy over existing methods.
• CNNs offer a powerful tool for automated and accurate pollen identification.