MiekevanVlaardingen/Image_based_trait_extraction

Dataset Card

Dataset Summary

This dataset contains side-view RGB images of Chenopodium quinoa plants exposed to drought and salinity stress, collected at the Netherlands Plant Eco-phenotyping Centre (NPEC). The repository also includes the scripts required to preprocess the raw images, train and apply a U-Net++ segmentation model, manually annotated images used for model training, and the trained segmentation model accompanying the associated publication.

The segmentation model classifies each image into the following plant organs:

• Panicle
• Stem
• Supporting stick
• Leaves (remaining plant pixels)

The resulting segmented images enable automated, organ-specific quantification of plant development over time for high-throughput, non-destructive phenotyping.

---

Dataset Details

Dataset Description

This repository contains:

• Preprocessed side-view RGB images from two greenhouse experiments (Exp1 and Exp2)
• Manually annotated segmentation masks
• Scripts for preprocessing, model training, and inference
• The trained U-Net++ segmentation model used in the accompanying publication

The dataset contains images from two quinoa genotypes:

• B – Bastille
• P – Pasto

Images were collected under three treatment conditions:

• Control (C)
• Drought (D)
• Salinity (S)

---

Dataset Card Authors

• Mieke van Vlaardingen
• Eibertus N. van Loo

---

Curated by

The image dataset was generated and curated at the Netherlands Plant Eco-phenotyping Centre (NPEC), Wageningen University & Research.

Image acquisition and dataset preparation were carried out by:

• Robert van Loo
• Viviana Jaramillo Roman
• Darío Bonaventura Roca Campos
• Mieke van Vlaardingen

---

Funding

This work was supported by the Dutch Top Sector programme through project LWV21013 – Novel diversity in emerging crops with gene editing by chemical mutagenesis and CRISPR.

The project was carried out in collaboration with Radicle Crops, Plantik Bioscience, Lekker Lupine, and the Netherlands Plant Eco-phenotyping Centre (NPEC). In-kind contributions were provided by the industrial partners.

---

License

Creative Commons Attribution 4.0 International (CC BY 4.0)

---

Associated Publication

This dataset accompanies the following publication:

van Vlaardingen, M., To, J., Jaramillo Roman, V., Roca Campos, D. B., Walmsley, A., Perez-Borroto, L. S., Trindade, L. M., & van Loo, E. N. (2026).

Image-based trait extraction of Chenopodium quinoa grown under salinity and drought stress.

Journal of Experimental Botany.

DOI: https://doi.org/10.1093/jxb/erag358

Publication: https://academic.oup.com/jxb/advance-article/doi/10.1093/jxb/erag358/8741239

---

Dataset Structure

The repository contains the following folders:

Annotated_images_exp1_and_2

Manually annotated RGB images and corresponding segmentation masks used for training the U-Net++ model. Annotations were created using AnyLabeling.

Images_exp1

Preprocessed side-view RGB images from greenhouse experiment 1.

Images_exp2

Preprocessed side-view RGB images from greenhouse experiment 2.

Trained_segmentation_model

The trained U-Net++ model accompanying the associated publication.

Scripts

• PreprocessQuinoa.py – preprocessing of raw NPEC RGB images.
• Training_and_applying_unet – scripts for training and inference.
• Environment – required Python environment and package information.

All images in this repository have already been preprocessed and are ready for model training or inference.

---

Image Naming Convention

Each filename contains the experimental metadata.

Example:

20221111_BI3.POT02_BC_H1.Control.1_1.png

Component	Description
20221111	Image acquisition date (YYYYMMDD)
BI3	Imaging station identifier
POT02	Unique pot identifier
B	Genotype (B = Bastille, P = Pasto)
C	Treatment (C = Control, D = Drought, S = Salinity)
H1	Harvest or measurement time point
Control.1_1	Additional experimental identifiers
.png	Image format

---

Dataset Creation

Curation Rationale

High-throughput plant phenotyping enables continuous, non-destructive monitoring of plant development throughout an experiment. This dataset was generated to facilitate automated phenotyping of quinoa subjected to drought and salinity stress using deep learning-based image segmentation.

The U-Net++ model enables automated extraction of individual plant organs, allowing downstream analyses such as organ-specific growth and temporal trait extraction

---

Source Data

Raw side-view RGB images were collected using the conveyor-based imaging platform at the Netherlands Plant Eco-phenotyping Centre (NPEC).

Images were subsequently preprocessed using PreprocessQuinoa.py before manual annotation and model training.

Complete details regarding the experimental design, image acquisition, preprocessing pipeline, segmentation model, and validation are provided in the accompanying publication.

---

Data Collection and Processing

The dataset was generated using the following workflow:

1. Image acquisition using the NPEC side-view RGB imaging platform.
2. Image preprocessing using PreprocessQuinoa.py.
3. Manual image annotation using AnyLabeling.
4. Training of a U-Net++ semantic segmentation model.
5. Application of the trained model to generate organ-specific segmentation masks and pixel counts.

---

Software

The repository contains all scripts required to reproduce the segmentation workflow, including preprocessing, model training, and inference.

Manual annotations were generated using AnyLabeling.

---

Personal and Sensitive Information

This dataset contains no personal, confidential, or sensitive information.

---

Citation

If you use this dataset in your research, please cite both the dataset and the accompanying publication.

BibTeX

@misc{QuinoaSegmentationDataset2026,
  title = {Quinoa image segmentation dataset for image based trait extraction under salinity and drought stress},
  author = {Mieke van Vlaardingen, Jacky To and Eibertus N. van Loo},
  year = {2026},
  howpublished = {\url{https://huggingface.co/datasets/MiekevanVlaardingen/Image_based_trait_extreaction}}
}

@article{vanvlaardingen2026image,
  author = {van Vlaardingen, Mieke and To, Jacky and Jaramillo Roman, Viviana and Roca Campos, Darío Bonaventura and Walmsley, Aimee and Perez-Borroto, Lucia Sandra and Trindade, Luisa M. and van Loo, Eibertus N.},
  title = {Image-based trait extraction of {Chenopodium quinoa} grown under salinity and drought stress},
  journal = {Journal of Experimental Botany},
  year = {2026},
  doi = {10.1093/jxb/erag358}
}

---

Repository Overview

Contents

Images

• Images_exp1
• Images_exp2

These folders contain the preprocessed side-view RGB images from the two greenhouse experiments.

Annotated Images

Annotated_images_exp1_and_2

Contains manually annotated RGB images and their corresponding segmentation masks used for training the U-Net++ model.

Trained Model

Trained_segmentation_model

Contains the trained segmentation model accompanying the publication.

Scripts

The repository includes scripts for preprocessing, training, and inference:

• Scripts/PreprocessQuinoa.py
• Scripts/Training_and_applying_unet
• Scripts/Environment

---

Training a New U-Net++ Model

All scripts required for training are located in

Scripts/Training_and_applying_unet

Main Files

config.py

Contains all user-defined settings, including:

• Training image directory
• Annotation directory
• Output directory
• Model configuration

inference.py

Applies a trained segmentation model to new images and can optionally evaluate segmentation performance.

launch_train.sh

Launches model training on a GPU.

---

Configuration

To train a new model, only modify the paths specified in config.py.

Update:

• Training image directory
• Annotation directory
• Output directory

No additional files normally require modification.

---

Running the Training

Start model training using

launch_train.sh

This script submits the training job to a GPU-enabled compute node.

---

Evaluating Model Performance

To evaluate the trained model on annotated images, set

TRACK_PERFORM = True

This compares predicted segmentation masks with the manually annotated ground-truth masks.

For inference on new images, set

TRACK_PERFORM = False

---

Applying a Trained Model

To apply an existing model:

1. Modify the required input, output, and model directories in inference.py.
2. Set

TRACK_PERFORM = False

3. Run

launch_apply2.sh

---

Output

The trained model produces:

• Organ-specific segmentation masks
• Pixel counts for each plant organ
• Segmented RGB images

These outputs can subsequently be used for downstream phenotypic analyses, including temporal growth analysis, organ-specific trait extraction, and statistical modelling using linear mixed-effects or other regression approaches.

---

Contact

For questions regarding the dataset or the accompanying publication, please contact:

Mieke van Vlaardingen Wageningen University & Research Email: mieke.vanvlaardingen@wur.nl

Robert van Loo Wageningen University & Research Email: robert.vanloo@wur.nl

Version

Current release: v1.0

This version corresponds to the dataset accompanying the publication:

Image-based trait extraction of Chenopodium quinoa grown under salinity and drought stress (Journal of Experimental Botany, 2026).