food-ai-nexus/salmonella-serovar-hyperspectral-spectra

Salmonella Serovar Hyperspectral Spectra (Foods 2025)

Salmonella Serovar Hyperspectral Spectra is a tabular dataset of single-cell spectral features for foodborne bacterial classification. It was created to support research in rapid pathogen identification, enabling models to classify Salmonella serovars using hyperspectral signatures extracted from individual bacterial cells.

Companion image dataset: The RGB composite microscopy images from which these spectra were extracted are available at food-ai-nexus/salmonella-serovar-hyperspectral.

This dataset accompanies the publication: Papa, M., Bhattacharya, S., Park, B., & Yi, J. (2025). Rapid Salmonella Serovar Classification Using AI-Enabled Hyperspectral Microscopy with Enhanced Data Preprocessing and Multimodal Fusion. Foods, 14(15), 2737. doi: 10.3390/foods14152737

Dataset Description

Each row represents one bacterial cell segmented from a hyperspectral data cube (hypercube). Spectra are Standard Normal Variate (SNV)-normalized mean single-cell spectra across 303 wavebands (399–1000 nm, 2 nm bandwidth), extracted using an attention-gated residual U-Net (ARG2U-Net).

from datasets import load_dataset

ds = load_dataset("food-ai-nexus/salmonella-serovar-hyperspectral-spectra")
# ds['train'] → 18,180 rows | ds['test'] → 7,792 rows

Splits

The 70/30 train/test split is performed at the row level, stratified by Serovar (seed=42), mirroring the paper's reported methodology.

Split	Rows	Notes
train	18,180	70% stratified by serovar
test	7,792	30% stratified by serovar

Schema

Column	Type	Description
InImage_ID	int	Per-serovar cell index identifying the source hypercube
Band_2_W_401.00 … Band_303_W_1000.90	float	SNV-normalized mean spectral reflectance at each waveband (nm)
Serovar	string	Target label: one of Enteritidis, I4, Infantis, Johannesburg, Kentucky

Important note on InImage_ID: This index identifies the source hypercube within each serovar group, not globally. It cannot be used as a direct foreign key to join rows to specific files in the companion image dataset.

Classes

Five Salmonella serovars selected based on their prevalence in foodborne illness outbreaks:

Label	Serovar	Train Rows	Test Rows
Enteritidis	S. Enteritidis	3,731	1,600
I4	S. 4,[5],12:i:-	3,638	1,559
Infantis	S. Infantis	5,201	2,229
Johannesburg	S. Johannesburg	3,265	1,399
Kentucky	S. Kentucky	2,345	1,005

Note on class imbalance: The spectra are inherently imbalanced because different serovars yield different numbers of segmentable cells per hypercube. This reflects biological variation in cell density and morphology, not a sampling artifact.

Known data issue: One cell record (InImage_ID=73, Enteritidis) has NaN values for bands 148–303 (wavelengths 692–1001 nm) in the original Zenodo source CSV. This row is preserved as-is to maintain source fidelity. Users should apply appropriate imputation or filtering before training.

Baseline Performance

Model	Modality	Test Accuracy
PCA-MLP	Spectral only	81.1%
PCA-MLP + EfficientNetV2	Multimodal fusion (Image + Spectra)	82.4%

License

This dataset is licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) License.

Citation

@article{papa2025salmonella,
  title   = {Rapid Salmonella Serovar Classification Using AI-Enabled Hyperspectral Microscopy with Enhanced Data Preprocessing and Multimodal Fusion},
  author  = {Papa, MeiLi and Bhattacharya, Siddhartha and Park, Bosoon and Yi, Jiyoon},
  journal = {Foods},
  volume  = {14},
  number  = {15},
  pages   = {2737},
  year    = {2025},
  doi     = {10.3390/foods14152737}
}

Source

Original dataset: Zenodo 10.5281/zenodo.16740800
Code repository: GitHub food-ai-engineering-lab/salmonella-serovar-classification-foods