HyperPRI: Hyperspectral Plant Root Imagery

28 Aug 2024 - Spencer J. Chang

Additional Authorship: R. Chowdhry, Y. Song, T. Mejia, A. Hampton, S. Kucharski, TM Sazzad, Y. Zhang, S. J. Koppal, C. H. Wilson, S. Gerber, B. Tillman, M. FR Resende Jr., W. M Hammond, A. Zare.

TL;DR - RGB, HSI, Temporal Rhizobox Dataset

  • Rhizobox dataset: Red-green-blue, hyperspectral, temporal, and mask annotation data for in-situ, belowground root imagery.
  • Sample features: High correlation, thin object features, and various imaged textures for soil and root regions.
  • Hyperspectral Imaging: Found to be useful for understanding plant drought responses belowground and for improving root-soil segmentation.
  • Paper, dataset, source code links are provided.
Figure 1: An RGB image from the HyperPRI dataset of peanut roots. The RGB image is on top, and the annotation mask is on bottom.

Why belowground drought responses?

To tackle future agricultural challenges of greater nutritional needs and hotter temperatures, researchers study high-yielding crop plants through phenotyping, part functions, and drought-resiliency experiments. In the area of “non-destructive” methods, researchers analyze plants using one of the following imaging perspectives:

The former fails to capture the belowground perspective, and X-ray CT and MRI technologies are too expensive to be deployed at scale into crop fields. RGB imaging underground focuses on physical root characteristics; it cannot be utilized to understand a plant’s response to drought conditions. On the other hand, HSI data of plant roots can give researchers insight into how plants respond to drought conditions.

Video 1: A demonstration of the large variety of spectral signatures in a single peanut plant root image. The recording was done on Windows using the TruScope software.

How does HSI data give more insight than RGB data?

The biggest difference comes down to the type of imaging sensor:

  • RGB – Visible Light Spectrum
    • Bandwidths ~635-700 nanometers (nm) (red), ~520-560 nm (green), and ~450-490 nm (blue).
    • 3 image layers for red, green, and blue colors.
  • HSI – Any wavelength in electromagnetic spectrum.
    • Examples: ultraviolet or infrared light.
    • Thin bandwidths but large range. Imaging a 2-nm width between 400 to 700 nm results in an “image cube” with 150 image layers (300 nm / 2 nm = 150).

Where is this useful compared to RGB data? Researchers can apply hyperspectral near-infrared (NIR) imaging to study the hyperspectral signatures (aka. spectral signatures) of plant leaves; this reveals chemical information related to a leaf’s water content.

  • Studies have investigated different plant phenotypic responses to drought conditions using each plant phenotype’s leaf spectral signatures (Trends in Plant Science 2022).
  • There’s growing evidence that spectral signatures can…
    • Characterize plants’ physiological properties
    • Provide an earlier indicator of plant stress under drier-than-normal conditions.

Most hyperspectral plant studies typically consider the aboveground perspective mentioned above. There is a tremendous lack of research done belowground, especially with HSI data. At the University of Florida, we have been collaborating with multiple labs to study how peanut (Arachis hypogaea) and sweet corn (Zea mays) plants respond to changes in water content below the soil surface. To do this, we grew the two species in plexiglass boxes (ie. rhizoboxes) and collected HSI data of plant roots for a timespan of two months under normal and drier-than-normal soil conditions.

Video 2: A brief introduction to the HyperPRI dataset.

What is the HyperPRI dataset?

Hyperspectral Plant Root Imagery (HyperPRI) is a dataset of about 745 images containing temporal, hyperspectral, and RGB data of peanut and sweet corn plant roots in rhizoboxes.

  • Data covers the visible-NIR (VIS-NIR) range of 400-1000 nm for an imaging area of 9.3 x 14.8 cm2.
    • Spectral resolution: 2 nm (299 bands)
    • Image resolution: 0.15 mm (968 x 608).
    • Rhizoboxes Planted: 32 peanut and 32 sweet corn
    • Includes box weights and fully annotated segmentation masks
  • It has imaging data for drought-level conditions.
    • Peanut plants: Occurs 75 - 86 days after planting
    • Sweet Corn plants: Occurs 39 - 62 days after planting
    • Some of each species were held out as control (always well-watered).
  • Each box was imaged on 14 or 15 days during two months of data collection (Figure 2 shows the rhizobox timeline).
Figure 2: A depiction of the growth timeline for the rhizobox plants. From left to right, the images show the dry soil, the wet (watered) soil, the watered plant, the plant in drought, and the rewatered plant.
Figure 3: Correlation between different HS bands for a single image.

What features does the dataset have?

In addition to the temporal aspects of the dataset, HyperPRI contains highly correlated, high-dimensional data (Figure 3). Here’s a quick rundown of those not already covered:

  • Figure 4: Root and soil show different spectral signature distributions.
  • Figure 5: Root thicknesses vary and can be extremely thin (1-3 pixels).
  • Figure 6: Diversity of visual textures and colors for both root and soil.
Figure 4: Root and soil spectral signature distributions for peanut root images. These statistics are drawn from a portion of pixels across approximately 15 peanut root images.
Figure 5: Sample images of peanut roots that are 1 – 3 pixels wide and in various soil conditions. Thin roots become less visible in certain soil conditions.
Figure 6: Sample images of visual textures for soil (top) and root (bottom) regions available in HyperPRI.

How can you use the dataset?

HyperPRI may be used in numerous ways:

  1. Compute physical root characteristics: (length, diameter, angle, count, system architecture)
  2. Analyze root turnover: root decay and growth
  3. Study drought resiliency and response: utilize box weight to indirectly link amount of water and plant growth across time.
  4. Follow physical and hyperspectral root traits over time: use the data at each time step to see how traits change.
  5. Compare physical and hyperspectral root traits between two crop plants: physical, hyperspectral, and temporal aspects can be compared.
  6. Investigate different texture features for roots and soil: soil has various conditions (dry, wet, condensation, etc.), root can be amid algae or mold and can show signs of decay.
  7. Segment roots from soil: Tackle a difficult task of generalizing segmentation across multiple watering conditions.

I have used approximately 60 peanut root images to cross-validate UNET-like models in the task of root-soil segmentation. Using three models, we demonstrate that HyperPRI includes hyperspectral information that improves segmentation performance. The three models are as follows:

  1. The spatial model uses the RGB root images like a UNET model to determine root-soil boundaries. This convolutional model focuses on the spatial information in each RGB image layer.
  2. The spectral model uses individual pixels’ spectral signatures to determine root-soil boundaries. In this perspective, each spectral signature is passed through a UNET-like MLP model; introduced in the paper as SpectralUNET, this model only learns hyperspectral features.
  3. The spatial-spectral model uses the root HSI data by replacing the UNET’s first 2D convolution layer with a 3D convolution layer spanning the entire spectral signature. In other words, the 2D convolution layer uses a learnable 3x3 filter, and with our chosen spectral bands, the 3D convolution layer uses a learnable 3x3x238 filter. Named CubeNET, this model simultaneously learns spatial and hyperspectral features.

When we cross-validate the three models, we notice that CubeNET generalizes better and predicts true negatives with greater consistency (see Figure 7). Comparing the different models’ segmentations, we note that CubeNET does better at segmenting roots in dry images and thin roots, although it tends to undersegment root pixels (Figure 8).

Figure 7: Test results from training UNET, SpectralUNET, and CubeNET on HyperPRI.
Figure 8: Test segmentation results for UNET, SpectralUNET, and CubeNET.

BibTeX Citation

@article{CHANG2024109307,
    title = {HyperPRI: A dataset of hyperspectral images for underground plant root study},
    journal = {Computers and Electronics in Agriculture},
    volume = {225},
    pages = {109307},
    year = {2024},
    issn = {0168-1699},
    doi = {https://doi.org/10.1016/j.compag.2024.109307},
    url = {https://www.sciencedirect.com/science/article/pii/S0168169924006987},
    author = {Spencer J. Chang and Ritesh Chowdhry and Yangyang Song and Tomas Mejia and Anna Hampton and Shelby Kucharski and T.M. Sazzad and Yuxuan Zhang and Sanjeev J. Koppal and Chris H. Wilson and Stefan Gerber and Barry Tillman and Marcio F.R. Resende and William M. Hammond and Alina Zare},
    keywords = {Root physiology, Hyperspectral imagery, Minirhizotron, Rhizotron, Semantic segmentation},
}

Where can you learn more?

If you’re interested in learning more, check out the following links: