Datasets:

AVLL
/

PGP_v2

A Data-Driven Image Extraction and Analysis Pipeline for Plant Phenotyping in Controlled Environments Fahimeh Orvati Nia, Joshua Peeples, et al. — The Plant Phenome Journal

This release contains the Plant Growth and Phenotyping version 2 (PGP v2) dataset collected in the Texas A&M AgriLife Automated Precision Phenotyping Greenhouse (APPG). PGP v2 expands the earlier PGP v1 release and includes longitudinal multispectral greenhouse imagery of maize, cotton, rice, and sorghum, together with processed pseudo-RGB images and annotation products used for computer-vision and plant-phenotyping analyses. It also includes an independent maize cold-stress RGB dataset (Case Study 2) used to evaluate the generalizability of the analysis pipeline beyond the APPG robotic imaging system.

Release status: the core/ bundle (pseudo-RGB images, segmentation masks, and COCO annotations for all four crops) is available now. The raw_multispectral/ bundle and the Case Study 2 (CERCA maize cold-stress) bundle are being added in follow-up updates — see the Files table below for per-file status.

See DATASET_MANIFEST.csv in this folder for the release-level inventory, package sizes, file counts, filename patterns, examples, and package-specific notes.

Dataset at a glance

The manuscript reports the following PGP v2 image collection:

Crop	Date range	Imaging sessions	Imaging frequency	Number of plants	Frames per plant	Total images
Maize	Dec 2023–Jul 2024	40	3–4×/week	271	6–15	14,448
Cotton	Apr 2024–Jul 2024	66	Daily	12	5–10	27,660
Sorghum	Dec 2024–May 2025	17	Weekly	48	13	10,608
Rice	May 2024–Aug 2024	8	Weekly	8	10–26	1,376

Total reported in the manuscript: 54,092 images.

What is in this dataset?

PGP v2 contains repeated plant-imaging sessions collected in a controlled greenhouse. Plants were imaged on fixed benches using a robotic gantry. During an imaging run, the camera moved through the greenhouse and captured vertically stacked sequences of overlapping frames from the plant base toward the canopy apex.

The dataset is intended to support tasks including:

plant detection and segmentation
semantic and instance segmentation
plant-instance tracking across vertically stacked frames
image stitching / plant mosaic reconstruction
vegetation-index analysis
texture analysis
morphological trait extraction
longitudinal plant phenotyping
development and evaluation of computer-vision and deep-learning methods

The full analysis pipeline used in the companion manuscript is available at: https://github.com/Advanced-Vision-and-Learning-Lab/Plant_Analysis_Tool_Pipeline

Project website: https://advanced-vision-and-learning-lab.github.io/PGP_Website/

Multispectral imaging system

APPG multispectral images were collected using the MSISAGRI1A multispectral camera (Spectral Devices Inc.). The camera uses a 4-megapixel CMOS sensor and captures four spectral bands simultaneously:

Band	Center wavelength	Description in the manuscript
Yellow / yellow-green	580 nm	Used as the surrogate visible/green channel in the analysis pipeline
Red	660 nm	Visible red band
Red-edge	735 nm	Red-edge spectral band
Near-infrared (NIR)	820 nm	Near-infrared spectral band

Each multispectral frame is stored as a 16-bit TIFF image at 512 × 512 pixels with the four spectral channels.

The imaging system used synchronized LED illumination and AntiXTalk technology to reduce inter-band leakage. Exposure time and gain were held fixed at 1000 µs and 1.0, respectively, and LED illumination settings were held constant across imaging sessions.

Important spectral-data limitation: the raw multispectral images were not converted to absolute reflectance. The manuscript states that no dark-reference correction, white-reference correction, sensor non-uniformity correction, or other radiometric calibration was applied before analysis. Therefore, spectral measurements in this release should be interpreted as image-derived spectral intensities, not absolute reflectance measurements.

Pseudo-RGB images

For compatibility with standard computer-vision models, the four-band multispectral images were converted to 8-bit pseudo-RGB composites. Each spectral band was independently min-max normalized to the 0–255 range and cast to unsigned 8-bit format. The pseudo-RGB composite uses:

580 nm yellow/yellow-green band as the surrogate green channel
735 nm red-edge band
660 nm red band

The 820 nm NIR band was retained separately for vegetation-index and spectral analysis.

This normalization is a processing step for visualization and model compatibility and does not constitute radiometric or color calibration across imaging sessions.

Case Study 2 — CERCA maize cold-stress dataset

This release also includes RGB images from an independent maize cold-stress experiment conducted at the University of Nebraska–Lincoln CERCA facility, used to evaluate the generalizability of the analysis pipeline beyond the APPG robotic imaging system. The dataset contains 50 RGB images from 10 maize genotypes, with one image per genotype under five conditions: before treatment, after 10 °C treatment, after 4 °C treatment, recovery following 10 °C treatment, and recovery following 4 °C treatment. Images were acquired using a Canon EOS M10 camera and are provided in Maize-CS-2-RGB-1.zip and Maize-CS-2-RGB-2.zip. The companion manuscript uses these images to evaluate RGB-derived vegetation-index, texture, and morphological features for characterizing variation associated with cold stress and recovery.

Image acquisition and sequence structure

Each plant is represented by a vertical sequence of frames captured at successive robot positions along the Z-axis. Taller plants require more frames to cover the complete plant profile.

For maize, cotton, and sorghum, the manuscript reports an approximate 50 mm vertical spacing between consecutive camera positions, with scans beginning near a robot Z-position of 925 mm. The upper scan position was selected according to crop and canopy height.

Consecutive frames intentionally overlap spatially so that they can be used for:

plant-instance association
frame-to-frame matching
image stitching
full-plant reconstruction
trait aggregation across views

A fixed numerical overlap percentage was not recorded in the acquisition metadata.

Recorded acquisition metadata

For each image acquisition, the APPG system records metadata including: species, room number, date, time, image index, row number, plant number, orientation, plant spatial coordinates, and robot spatial coordinates. These metadata support plant identification, spatial calibration, and longitudinal tracking.

Treatment and genotype annotations are not uniformly available for every component of PGP v2. Where available, those annotations are included in the associated metadata.

Release files

The release is organized into three logical bundle types.

Core bundles

Core bundles contain processed imagery and annotation products intended for most computer-vision users.

File	Crop	Contents	Manifest item count	Approx. size	Status
`core/rice_core.zip`	Rice	pseudo-RGB images + semantic masks + instance masks + COCO annotations	2,791	0.28 GB	✅ available
`core/sorghum_core.zip`	Sorghum	pseudo-RGB images + semantic masks + instance masks + COCO annotations	17,609	3.85 GB	✅ available
`core/maize_core.zip`	Maize	pseudo-RGB images + semantic masks + instance masks + COCO annotations	19,864	5.11 GB	✅ available
`core/cotton_core.zip`	Cotton	pseudo-RGB images + semantic masks + instance masks + COCO annotations	71,648	10.08 GB	✅ available

Important: the core-bundle item count is the number of packaged files/items and should not be interpreted as the number of unique raw plant images, because each bundle can contain images, masks, and annotation files.

Raw multispectral bundles

Raw multispectral bundles contain the original four-band 16-bit TIFF frames.

File	Crop	Contents	Raw frame count in release manifest	Approx. size	Status
`raw_multispectral/rice_multispectral.zip`	Rice	4-band 16-bit TIFF frames	1,376	1.31 GB	✅ available
`raw_multispectral/sorghum_multispectral.zip`	Sorghum	4-band 16-bit TIFF frames	10,607	10.40 GB	✅ available
`raw_multispectral/maize_multispectral.zip`	Maize	4-band 16-bit TIFF frames	14,448	16.22 GB	✅ available
`raw_multispectral/cotton_multispectral.zip`	Cotton	4-band 16-bit TIFF frames	27,660	31.62 GB	✅ available

Case Study 2 bundle

File	Crop	Contents	Item count	Status
`case_study_2/Maize-CS-2-RGB-1.zip`	Maize (CERCA cold-stress)	RGB images, Canon EOS M10	part of 50 total	✅ available
`case_study_2/Maize-CS-2-RGB-2.zip`	Maize (CERCA cold-stress)	RGB images, Canon EOS M10	part of 50 total	✅ available

Core bundles total: 19.29 GB | Raw multispectral total: 59.55 GB | Grand total (core + raw multispectral): 78.83 GB | Case Study 2: to be added

Which files do I need?

For most computer-vision and segmentation work, download the core/ bundle for the crop(s) of interest. Use the raw_multispectral/ bundle when you need the original four-band 16-bit TIFF data for spectral analysis, vegetation-index development, alternative pseudo-color construction, custom normalization or preprocessing, or methods that operate directly on multispectral channels. If you need both processed products and the original spectral data, download both the core and raw multispectral bundles for that crop. Use the case_study_2/ bundle if you are evaluating pipeline generalizability on an independent RGB dataset with a cold-stress treatment design.

Filename conventions

Filename patterns differ slightly across crops.

Rice core: crop-date-plantN-frameN.ext Example: rice-2024-05-29-plant1-frame1.png

Sorghum core and raw multispectral: crop-date-plantN-frameN.ext Examples: sorghum-2024-12-04-plant10-frame1.png, sorghum-2024-12-04-plant10-frame1.tif

Maize and cotton: crop-date-rowN-cameraN-frameN.ext Examples: maize-2024-06-24-row1-camera2-frame21.png, maize-2024-06-24-row1-camera2-frame21.tif, cotton-2024-06-24-row1-camera2-frame21.png, cotton-2024-06-24-row1-camera2-frame21.tif

Rice raw multispectral: crop-date-cameraN-plantN-frameN.ext Example: rice-2024-05-29-camera1-plant1-frame1.tif

Data organization and intended reuse

PGP v2 was designed to support longitudinal and multi-crop controlled-environment phenotyping. Data are organized around crop identity, imaging date, plant identity, and frame position.

Potential reuse includes: benchmarking plant segmentation algorithms, training segmentation and detection models, testing instance-tracking methods, evaluating image-stitching algorithms, extracting plant morphology and canopy traits, computing vegetation indices, studying spectral and texture features, developing self-supervised learning methods, and longitudinal analysis of plant growth and stress.

The manuscript reports that the full processing framework generated 863 quantitative features per plant and imaging date, spanning vegetation indices, spectral statistics, texture descriptors, and morphological traits. These derived features are products of the companion analysis pipeline rather than a definition of every file included in the release package.

Hosting plan

Hugging Face Hub (fahimehorvatinia/PGP_v2) — primary host, for the citable DOI and for practical browsing / load_dataset(...) access; handles the full file/byte counts comfortably.
Link the permanent host from the existing project site: https://advanced-vision-and-learning-lab.github.io/PGP_Website/

Related resources

Project website: https://advanced-vision-and-learning-lab.github.io/PGP_Website/
Analysis pipeline: https://github.com/Advanced-Vision-and-Learning-Lab/Plant_Analysis_Tool_Pipeline

Citation

Please cite the companion article when using this dataset:

Orvati Nia, F., Peeples, J., et al. A Data-Driven Image Extraction and Analysis Pipeline for Plant Phenotyping in Controlled Environments. The Plant Phenome Journal.

Pipeline code: https://github.com/Advanced-Vision-and-Learning-Lab/Plant_Analysis_Tool_Pipeline

Downloads last month: 15