Million-scale multimodal pollen microscopy with expert-guided foundation models

2026-06-16 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, quick

Summary

Pollen AI Atlas is a million-scale multimodal pollen microscopy resource designed to overcome bottlenecks in automated pollen identification for aerobiology, palaeoecology, and biodiversity monitoring. This atlas comprises pure-species whole-slide bright-field images from four geographic origins, four scanner settings, and 46 taxon labels across 31 botanical families. Through token-level mining and filtering, seeded by one manual exemplar per slide, it generated 1,511,390 grain detections with 99.6% proposal precision. Each detection includes machine-generated grain-level morphological captions from five open-weight vision-language models, guided by expert-verified palynological anchors, detailing aperture systems, wall ornamentation, shape, and size. Gemma4 performed best for caption generation. Baseline benchmarks achieved 88.16% top-1 accuracy, and caption-derived text embeddings maintained robustness in cross-regional retrieval (mAP@20 0.811) even when image similarity degraded. The project releases data, annotations, captions, code, and weights.

Key takeaway

For AI Scientists and Computer Vision Engineers developing automated microscopy systems, the Pollen AI Atlas provides a critical, million-scale multimodal resource. You should utilize its 1.5 million expert-curated grain detections and machine-generated morphological captions to train and benchmark models for robust pollen identification across diverse conditions. This resource, including released data, code, and weights, offers a strong foundation for advancing cross-regional domain adaptation and domain-specific multimodal learning, reducing the bottleneck in aerobiology and palaeoecology.

Key insights

A million-scale multimodal pollen atlas uses expert-guided vision-language models to automate identification and provide structured morphological descriptions.

Principles

• Multimodal data improves robustness across diverse conditions.
• Expert guidance enhances VLM caption accuracy.
• Text embeddings maintain robustness despite image degradation.

Method

Seed with one manual exemplar per slide, perform token-level mining, filter detections, then generate captions using VLM guided by palynological anchors.

In practice

• Benchmark pollen recognition systems.
• Facilitate cross-regional domain adaptation.
• Advance domain-specific multimodal microscopy learning.

Topics

• Multimodal AI
• Pollen Identification
• Vision-Language Models
• Microscopy
• Domain Adaptation
• Computer Vision

Best for: AI Scientist, Computer Vision Engineer, Research Scientist

Related on AIssential

• A multi-modal dataset for insect biodiversity with imagery and DNA at the trap and individual level — MassID45 integrates bulk and individual insect imagery with DNA data to train high-throughput taxonomic classifiers.
• AI to help researchers see the bigger picture in cell biology — An AI framework disentangles shared and unique cellular data across measurement modalities for a holistic cell view.
• Tomato Multi-Angle Multi-Pose Dataset for Fine-Grained Phenotyping — TomatoMAP dataset enables accurate, reproducible fine-grained tomato phenotyping using multi-angle imagery and AI models.
• MAVEN: A Multi-stage Agentic Annotation Pipeline for Video Reasoning Tasks — MAVEN's agentic, multi-stage pipeline automates high-quality, domain-adaptable video annotation for VLM training.
• FruitEnsemble: MLLM-Guided Arbitration for Heterogeneous ensemble in Fine-Grained Fruit Recognition — A two-stage dynamic inference framework combines heterogeneous ensembles with MLLM-guided arbitration for fine-grained fruit recognition.
• ChartNet: A Million-Scale, High-Quality Multimodal Dataset for Robust Chart Understanding — Code-guided synthetic data generation significantly improves VLM chart understanding and reasoning capabilities.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.