Seeds vary strongly across species, but also within species because of maturity, storage, lighting and viewing angle. A 2024 Scientific Data Data Descriptor introduced LZUPSD, an RGB dataset for fine-grained seed recognition.
88 categories and 4,496 images
The curated dataset contains 4,496 images across 88 seed categories, with roughly 50 images per category. It was divided 4:1 into 3,625 training and 871 test images.
Smartphone-based acquisition
Images were captured with a Xiaomi Mix2 smartphone and macro lens in a small studio against black fabric. Lighting, seed angle and position were varied, and processed images were cropped to 192×272 pixels.
This makes the dataset relatively inexpensive to reproduce compared with hyperspectral acquisition.
High benchmark accuracy does not equal field accuracy
ResNet50 reached 93.2% accuracy and SENet 95.1% on the controlled closed-set task. Real seed inspection can contain soil, debris, damaged seeds and unknown species, so those benchmark numbers should not be transferred directly to field deployment.
The dataset is visually interesting as a seed-morphology collection, but its main purpose is computer-vision training and benchmarking. In 2025, paired RGB+hyperspectral seed datasets such as BiSID-5k extended this direction into spectral classification.
For related context, see Predicting Individual Chinese Cabbage Weight by Drone: R² > 0.72 Even 53 Days Before Harvest.
For related context, see How Much “Breeding” Happened in This AI Breeding Study? Reading a Synthetic-Data bioRxiv Proof of Concept.
For related context, see SW14 Rebalances Soybean Seed Weight, Protein and Oil by Disrupting a LEC1-Containing NF-Y Complex.
References
- Yuan M et al. A dataset for fine-grained seed recognition. Scientific Data. 2024;11:344. https://doi.org/10.1038/s41597-024-03176-5
- A bimodal image dataset for seed classification from the visible and near-infrared spectrum. Scientific Data. 2025. https://doi.org/10.1038/s41597-025-05979-6
Comments