Stomata Detection Images Across Angiosperms Using Deep Learning

431 micrographs of leaf prints from 19 plant species were used to train a deep learning model for stomata detection. The VGG19 architecture achieved an average F-score of 0.87 on the training set and 0.67 on an unseen test set of 595 images from 16 additional species. The dataset and workflow were developed by Sofie Meeus of Meise Botanic Garden to automate the collection of stomatal traits.

Use Cases

• Training stomata detection models based on the provided 431 micrographs.
• Benchmarking computer vision architectures for botanical image analysis based on the described VGG19, shallow, and deep models.
• Studying stomatal trait variation across angiosperm phylogeny based on the multi-species collection.
• Developing standardized protocols for stomatal data collection based on the described nail polish method.

Strengths

• Dataset includes images from 35 total species across the angiosperm phylogeny, suggesting taxonomic breadth.
• The best-performing model achieved 94% accuracy for stomatal counts on unseen images of species used in training.
• Methodology is described as an 'easy-to-use and adaptable workflow' based on state-of-the-art deep neural networks.

Limitations

• Row count and dataset size are unknown, which may limit suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.
• Data may reflect taxonomic bias inherent to the 35 sampled angiosperm species.

Provenance

Source

Meise Botanic Garden

Collection Method

Images were taken from leaf prints made using the nail polish method from herbarium specimens.