Dataset

The Astro Cultivators dataset combines RGB, depth, and hyperspectral imaging to monitor soybean growth and drought stress in controlled chamber experiments.

Dataset Overview

The project uses multiple imaging sources so plant health can be studied from more than one view. RGB images capture visible color and shape, depth images support structure and height related measurements, and hyperspectral images capture wavelength based plant response.

3

Image Modalities

RGB, depth, and hyperspectral data are used together to support plant stress analysis.

51

HS Bands

Hyperspectral TIFF files were filtered to keep valid 51 band images for consistent analysis.

6 AM to 1 PM

Main Light Window

Most processing focused on the grow light window so image quality and lighting were more consistent.

Sample Image Modalities

These examples show the three main types of imaging data used in the project.

RGB Image

Used for visible plant color, segmentation, VARI, area, and growth related traits.

Depth Image

Used to support structure based features such as height, depth variation, and canopy shape.

Hyperspectral sample image of soybean plants

Hyperspectral Image

Used to compute vegetation indices and capture plant responses beyond normal RGB color.

Imaging Modalities

RGB Images

RGB images from the Intel RealSense camera were used to segment visible plant regions and compute visible growth features such as area, height, and VARI.

Depth Images

Raw depth data was used to explore plant structure features such as depth standard deviation, surface roughness, canopy volume, and vertical spread.

Hyperspectral Images

Hyperspectral TIFF cubes were used to measure plant response across wavelength bands and compute vegetation indices such as NDVI, GNDVI, red edge NDVI, chlorophyll red edge, and simple ratio.

Experiment Sets

The dataset includes calibration work and drought stress experiments. The experiment log includes chamber settings, sensor use, watering schedules, plant dates, and notes about data quality.

Plant Set	Sensor Type	Lighting	Experiment Type	Notes
Calibration Cycle 1	Hyperspectral	Night mode setup	Camera setup and calibration	Messy early data with gaps. Not used for stress detection.
Calibration Cycle 2	Hyperspectral	Night mode setup	Cleaner calibration data	Used for hyperspectral preprocessing and vegetation index feature engineering.
Set 1	Hyperspectral and RealSense	7 hour photoperiod	Drought experiment	Included control and drought plants. RealSense data had timing drift and quality issues.
Set 2	Hyperspectral and RealSense	6 AM to 1 PM	Best dataset for final modeling	Included one control plant on the left and one drought plant on the right. Used for final RealSense stress detection modeling.
Set 3	Hyperspectral and RealSense expected	6 AM to 4 PM	Current or incomplete	Still needed clarification and complete data before final analysis.

Set 2 Dataset

Set 2 was the strongest dataset for final RealSense analysis because it had a cleaner imaging window and clear control versus drought setup.

265

Usable Frames

Usable RealSense frames were identified after manual review, filtering, and segmentation checks.

20

Days

The final RealSense workflow used daily aggregated data across 20 days.

2

Treatments

The left plant was control and the right plant was drought.

Hyperspectral Dataset Details

The hyperspectral preprocessing work cleaned raw TIFF files, validated 51 band cube structure, removed invalid frames, created NDVI masks, and saved vegetation related metadata.

TIFF cubes 51 band validation Grow light filtering NDVI masking Vegetation indices

RealSense Dataset Details

The RealSense workflow used RGB and raw depth images from the Intel RealSense D435 camera. RGB images supported segmentation and color based features, while depth images supported structure based plant traits.

RGB images Raw depth images Grounded SAM 2 Control plant Drought plant

Dataset Limitations

Some images were removed because they were outside the lights on window, corrupted, too dim, or not usable for feature extraction.
Some hyperspectral data was useful for preprocessing and feature extraction but was not fully fused into the final RealSense classifier.
Set 2 was the strongest dataset, but the final model still needs validation on more plant sets.
Intermediate manual measurements such as ruler based height and leaf area were not fully available for all image dates.

Data Infrastructure

The notebooks use Python, NumPy, Pandas, tifffile, OpenCV, Matplotlib, SciPy, scikit image, PyTorch, Grounded SAM 2, SAM 2.1, and scikit learn to organize files, load images, filter invalid data, calculate features, and export CSV outputs.

← Back to Home