A High-Resolution Multifocal RGB Pollen Grain Image Dataset for Deep Learning Computer Vision Tasks from Biobío...

Area of Science:

• Palynology
• Computer Vision
• Biodiversity Informatics

Background:

• Existing pollen datasets often lack taxonomic diversity, sample size, and high-resolution, pixel-accurate annotations.
• Accurate pollen identification is crucial for ecological studies, biodiversity monitoring, and understanding plant-pollinator interactions.

Purpose of the Study:

• To introduce PollenBB16, a novel RGB pollen image dataset featuring pixel-accurate instance segmentation masks for Chilean flora.
• To provide a high-quality, taxonomically verified resource to address limitations in current palynological datasets.
• To facilitate advanced deep learning models for pollen classification and enable interdisciplinary research.

Main Methods:

• Collected 16,198 high-resolution brightfield microscopy images of 16 plant species from Chile's Biobío Region.
• Generated 36,383 pixel-accurate polygons with expert palynologist verification for taxonomic reliability.
• Captured images at three focal planes to reveal exine and internal grain structures, enabling robust feature extraction.

Main Results:

• The PollenBB16 dataset offers rich, multi-focal information for training more accurate convolutional neural networks.
• A baseline YOLOv11-seg model achieved a mask mAP@50 of 0.985, demonstrating the dataset's utility and establishing a performance benchmark.
• A leakage-safe data partition ensures the integrity of training and validation sets.

Conclusions:

• PollenBB16 significantly advances the field of automated pollen analysis by providing a comprehensive, high-quality dataset.
• The dataset supports diverse applications including aerobiology, climate change impact assessment, ecological restoration, and authentication of Chilean honey origins.
• This resource empowers interdisciplinary research bridging palynology, machine learning, and environmental science.