Faehn et al.

Plant Methods (2024) 20:171 Plant Methods

https://doi.org/10.1186/s13007-024-01297-x

METHODOLOGY Open Access

Advancing hyperspectral imaging

techniques for root systems: a new pipeline
for macro‑ and microscale image acquisition
and classification
Corine Faehn1*, Grzegorz Konert1,2, Markku Keinänen1,3,4, Katja Karppinen1,5 and Kirsten Krause1,5

Abstract
Background Understanding the environmental impacts on root growth and root health is essential for effective
agricultural and environmental management. Hyperspectral imaging (HSI) technology provides a non-destructive
method for detailed analysis and monitoring of plant tissues and organ development, but unfortunately examples
for its application to root systems and the root-soil interface are very scarce. There is also a notable lack of standard-
ized guidelines for image acquisition and data analysis pipelines.
Methods This study investigated HSI techniques for analyzing rhizobox-grown root systems across various imaging
configurations, from the macro- to micro-scale, using the imec VNIR SNAPSCAN camera. Focusing on three graminoid
species with different root architectures allowed us to evaluate the influence of key image acquisition parameters
and data processing techniques on the differentiation of root, soil, and root-soil interface/rhizosheath spectral signa-
tures. We compared two image classification methods, Spectral Angle Mapper (SAM) and K-Means clustering, and two
machine learning approaches, Random Forest (RF) and Support Vector Machine (SVM), to assess their efficiency
in automating root system image classification.
Results Our study demonstrated that training a RF model using SAM classifications, coupled with wavelength
reduction using the second derivative spectra with Savitzky-Golay (SG) smoothing, provided reliable classifica-
tion between root, soil, and the root-soil interface, achieving 88–91% accuracy across all configurations and scales.
Although the root-soil interface was not clearly resolved, it helped to improve the distinction between root and soil
classes. This approach effectively highlighted spectral differences resulting from the different configurations, image
acquisition settings, and among the three species. Utilizing this classification method can facilitate the monitoring
of root biomass and future work investigating root adaptations to harsh environmental conditions.
Conclusions Our study addressed the key challenges in HSI acquisition and data processing for root system analysis
and lays the groundwork for further exploration of VNIR HSI application across various scales of root system studies.
This work provides a full data analysis pipeline that can be utilized as an online Python-based tool for the semi-auto-
mated analysis of root-soil HSI data.
Keywords Graminoid, Root phenotyping, VNIR SNAPSCAN, Image analysis, Biomass estimation, Root-soil interface

*Correspondence:
Corine Faehn
[email protected]
Full list of author information is available at the end of the article

© The Author(s) 2024. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which
permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the
original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or
other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line
to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory
regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this
licence, visit http://creativecommons.org/licenses/by/4.0/.
Faehn et al. Plant Methods (2024) 20:171 Page 2 of 17

Introduction that predict root and soil water potentials, enhancing

Root-soil interactions in the rhizosphere are vital for our understanding of drought tolerance and recovery in
plant and ecosystem health. The rhizosheath, where root crops [18, 19]. The availability of such data, along with
exudates cause soil particles to adhere, plays a key role detailed acquisition methodologies and spectral signa-
in these interactions by enhancing root water retention tures, is crucial for advancing research on rhizosphere
and plant resilience against environmental stresses [1]. processes. When integrated with other analytical tech-
Robust root systems, together with their rhizosheath, niques, such as physiological phenotyping and functional
improve water and nutrient uptake, stabilize soil struc- genomics, HSI can be a powerful tool to complement the
tures, and prevent soil erosion [2, 3], while also protect- genotype-to-phenotype gap as part of a comprehensive
ing plants against soil-borne diseases [4]. Despite their research approach [20, 21].
importance, studying these interactions poses significant While traditional HSI systems, such as linescan or
challenges due to the hidden and complex nature of root pushbroom, rely on mechanical scanning, requiring lin-
systems below the soil surface [5]. ear movement of either the object or the camera to cap-
The use of rhizoboxes or rhizotrons—thin soil-filled ture the complete and spectral range, snapshot cameras
chambers with transparent observation windows—has capture the entire field of view without the need for spa-
enabled non-destructive surveillance and imaging of root tial scanning. The SNAPSCAN camera (imec, Leuven,
development, providing valuable insight into key root Belgium) merges linescan and snapshot imaging prin-
responses to different soil and environmental conditions ciples using on-chip filter technology, which simplifies
over time [6–8]. Innovative methods have been devel- the system assembly and enhances its usability for root
oped to integrate these set-ups with luminescence-based phenotyping. This camera, adaptable for use with front
reporters allowing for the examination of root architec- optics or microscope integration, has shown promise
ture and gene expression in soil-grown roots [9]. Addi- in various agricultural contexts, including plant species
tionally, the use of transparent tubes equipped with classification [22], estimating fruit maturity [23], and out-
cameras can be inserted into the soil, enabling the cap- door weed detection [24]. Its application in microscopy
ture of 360° images for in-situ monitoring of root devel- has primarily been in biomedical contexts [25–27], but to
opment in the field [10, 11]. Hyperspectral imaging (HSI) our knowledge, it has not yet been used in root pheno-
offers new possibilities for studying these complex inter- typing. Despite these valuable advances, numerous chal-
actions in greater detail. HSI captures a broad spectrum lenges must be addressed before the utilization of HSI to
of electromagnetic radiation beyond the visible range, analyze root systems will be comparable to its application
providing unique spectral signatures for each pixel. This in other areas of plant research.
technology not only extends our visual perception but Although the aforementioned articles have dem-
can also provide qualitative and quantitative information onstrated the SNAPSCAN camera’s versatility, the
on the physiological state and chemical composition of quality of the HSI data depends on user-selected con-
roots and the surrounding soil without extensive chemi- figurations and acquisition settings, a critical aspect
cal analyses [12]. that has received limited attention in the existing litera-
The use of HSI for root phenotyping of soil-grown ture. Manually adjusted settings, including the distance
plants is a relatively new approach. Traditionally, many between the sample and lens, lens aperture, and critical
HSI approaches have focused on aboveground data to software parameters such as time delay integration (TDI)
indirectly assess root health status [13, 14], but recent pixel step, pixel binning, and integration time, affect the
studies have expanded its direct application to root duration of image acquisition, spatial resolution, and the
imaging. Bodner et al. [12] demonstrated HSI’s capa- signal-to-noise ratio. These factors are essential because
bility to detect the radial composition and decomposi- they directly influence the quality of the acquired image
tion dynamics of root axes using spectral signatures in data. Various settings can be adjusted to balance between
the 1000–1700 nm range, which can be combined with acquiring high-quality images and faster image capture,
RGB imaging to determine root structural traits [5]. however, the subsequent image processing steps are also
Additionally, VNIR HSI has been utilized to predict lead fundamental for refining the data. These steps typically
stress levels in oilseed rape leaves and roots [15], clas- involve eliminating dead pixels, selecting specific regions
sify growth years of Kudzu roots [16], and distinguish of interest (ROI), enhancing spectral features through
between leaf mold and soil in the rhizosphere [17]. VNIR pre-processing, and compressing the image to retain
HSI has also recently been used to monitor the roots of only pertinent information [28]. While the SNAPSCAN
peanut and sweet corn under varying drought conditions, software automatically performs some pre-processing
with this data available in a publicly accessible HyperPRI steps, enabling immediate exploration of the data, the
dataset [18]. This dataset was useful to develop models choice of further data processing is objective dependent.
Faehn et al. Plant Methods (2024) 20:171 Page 3 of 17

Consequently, HSI data processing often necessitates plexiglass frame (2.5 cm × 20 cm × 0.6 cm) in between
tailored solutions adapted to the specific experimental giving a spatial volume of 247.5 ­cm3 (Fig. 1A). The back
settings. panel, two side frames, and bottom frame were glued
To effectively utilize HSI data for investigating root together before the rhizoboxes were filled with pre-mois-
systems and the biochemical composition in root-soil tened peat soil (Fig. 1B). The roots of all graminoids were
interactions, it is crucial to fully comprehend the capa- cut to approximately 4 cm in length and transplanted
bilities and limitations of this technology. Therefore, this individually in rhizoboxes at a depth of 1 cm below the
study aimed to investigate the techniques of image acqui- soil surface (Fig. 1C). Then the front panel was fastened
sition and data processing for evaluating plant root sys- with screws and hex nuts allowing easy removal of the
tems using VNIR SNAPSCAN technology across three front panel for later imaging. The rhizoboxes were placed
distinct dimensional scales: from an overview scale that in opaque plastic bags to block light entry and positioned
captures the entire rhizobox to a microscopic scale focus- at a 45-degree angle, with the front panel facing down to
ing on individual roots using a stereomicroscope. Since promote root growth along this imaging plane. Through-
different plant root traits such as diameter, density, and out the experiment, the rhizoboxes were kept in the same
rhizosheath composition may present unique challenges greenhouse conditions described above and watered fre-
for HSI, we selected three graminoid species with dis- quently to maintain high moisture levels.
tinct root system characteristics: Deschampsia flexuosa,
Eriophorum vaginatum, and Anthoxanthum odoratum. HSI system set‑up and acquisition parameters
All three species are well-adapted to survival in nutrient- Hyperspectral imaging (HSI) was performed using the
poor, acidic soils, and the low temperatures of subarctic VNIR SNAPSCAN camera (imec, Leuven, Belgium).
ecosystems, but have different root growth strategies The SNAPSCAN sensor has a spectral resolution of 150
[29–31]. Our analysis assesses whether the SNAPSCAN bands in the 470–900 nm wavelength range, and a spatial
camera can distinguish root traits of different species
across varied imaging scales to explore the applicability of
HSI in studying root adaptations to harsh environmental
conditions. We employed a methodical strategy that uti-
lized a small set of samples to distinguish between roots,
soil, and the root-soil interface by varying image acquisi-
tion settings and evaluating the data through image clas-
sification and processing techniques. Additionally, we
discuss potential challenges associated with the use of
the SNAPSCAN camera and provide recommendations
regarding the technical framework for future experimen-
tal set-ups focused on analyzing root-soil interactions.

Materials and methods

Plant material and rhizobox cultivation
Whole, intact plants including the root systems of the
grass D. flexuosa and the sedge E. vaginatum were col-
lected from a natural peat bog at Håkøybotn, Tromsø,
Norway (69° 63’N, 18° 78’E) in late summer of 2021. The
grass A. odoratum was collected from a previously reveg-
Fig. 1 Experimental set-up for root growth and image acquisition
etated urban site at Holt, Tromsø, Norway (69° 65’N, configurations. A Empty rhizobox, B rhizobox pre-filled with peat
18° 91’E). D. flexuosa and A. odoratum are true grasses, soil and (C) opened rhizobox with root system on the soil
which have fibrous and highly branched perennial root surface. D Configuration 1 (CONF1) for SNAPSCAN VNIR imaging
systems, while E. vaginatum has thick, unbranched including the camera equipped with a Schneider Kreuznach
annual root systems [30]. All plants were propagated veg- Apo-Xenoplan lens mounted on a frame connected to the imaging
stand with a 34.5 cm working distance (WD) between the camera
etatively in a greenhouse (15 °C, 18 h light, and photosyn- lens and the sample surface. E CONF2 using the same set-up
thetic photon flux density (PPFD) of 200 μmol ­m–2 ­s–1) at as CONF1 with a WD of 14 cm. F CONF3 used the camera with a 0.5x
the Climate Laboratory in Holt, Tromsø, Norway. C-mount lens adapter mounted to a stereomicroscope. The WD
Individual rhizoboxes consisted of two clear plexi- between the stereomicroscope lens and sample surface varied
glass panels (20 cm × 30 cm × 0.15 cm), two plexiglass between 10 and 14 cm according to the different magnifications
side frames (2.5 cm × 27.5 cm × 0.6 cm) and a bottom
 Faehn et al. Plant Methods

Table 1 Acquisition parameters tested for each of the three configurations for the VNIR SNAPSCAN camera
(2024) 20:171

Settings # images # images used to train Total #

used for pre- machine learning images for
classification model classification
CONF Lens WD CS CD Bin TDI MPB Lens IT SAM K-Means
aperture /
magnification

1 Schneider Kreuznach Apo-Xenoplan 34.5 cm 764 × 1024 115 × 155 mm 2 × 2 3 2.5 f/2.8 4–7 ms – – – 18
lens f/4.0 10–15 ms – – – 18
f/5.6 15–20 ms 12 12 6 30
2 Schneider Kreuznach Apo-Xenoplan 14 cm 1528 × 2048 49 × 66 mm 1×1 1 0 f/8.0 18 ms 6 6 – 8
lens f/11.0 18 ms 9 9 9 16
2.5 f/11.0 18 ms 3 3 – 3
5 0 f/11.0 18 ms 3 3 – 3
3 0.5X lens + stereomicroscope ~ 14 cm 800 × 800 15 × 15 mm 1×1 1 0 0.63X 10 ms 9 9 9 9
~ 11 cm 8 × 8 mm 1.6X 10 ms 9 – – 9
~ 10 cm 5 × 5 mm 4X 30 ms 9 – – 9
Total images: 60 42 24 123
CONF configuration, WD working distance, CS cube shape, CD cube dimension, Bin. binning, TDI time delay integration, MPB max. pixel blur, IT integration time, SAM spectral angle mapper
Page 4 of 17
Faehn et al. Plant Methods (2024) 20:171 Page 5 of 17

resolution of up to 3650 × 2048 pixels (7 Mpixels RAW datacube import, pixels with a reflectance value above 1
per band). The sensor frame rate has a maximum of 340 (overexposed) were set to 0. K-Means clustering was run
fps. Four halogen lamps (2000 K) equipped with diffusers with three clusters and a maximum of 20 iterations. Each
and 11.83 V and 6.72 A of power were used for illumi- classification method gave a classified image with three
nation. Lamps were connected to the imaging stand pro- classes and the corresponding average spectral reflec-
vided by imec, consisting of a viewing stage and frame to tance values for each class.
hold the camera and lamps. Hyperspectral images were The spectral data for each of the classification meth-
taken at three different configurations (CONFs), where ods were imported using the pandas v1.4.2 module [33].
CONF1 (Fig. 1D) and CONF2 (Fig. 1E) utilized the Sch- To reduce the large number of wavelengths to a smaller
neider Kreuznach Apo-Xenoplan lens (f2.0, C-mount, number of representative variables, Savitzky-Golay (SG)
focal length 24 mm, provided by imec) with different dis- smoothing [34] using the second derivative was applied
tances between the sample and lens, and CONF3 (Fig. 1F) using the SciPy v1.7.3 module [35]. The second derivative
utilized a 0.5x C-mount lens adapter in the ocular of a emphasizes small spectral variations and removes some
Leica MS5 stereomicroscope (Leica Microsystems, Wet- residual scattering effects, mainly additive effects and lin-
zlar, Germany), used with 0.63x, 1.6x, and 4x magnifica- ear baseline shifts [36], thus facilitating the selection of
tion. The front panel of the rhizobox was removed for the most informative bands from the spectrum. A com-
imaging in every configuration to avoid transmittance parison of window sizes was applied to the second deriv-
effects of the plexiglass panel. The acquisition parameter ative of the root spectra data to reduce the effect of noise
used for each configuration are listed in Table 1. but maintain the most important spectral information.
The hyperspectral images were collected using the Using the chosen window size, the average of the second
HSI SNAPSCAN (V1.8.1.1) software. The white refer- derivative was taken across each root spectra data set to
ence image was acquired by scanning the white reference identify the most informative bands.
target (provided by imec) at the settings used for sample
acquisition with reflectance set to 95%. The dark refer- Implementation of machine learning for image
ence image was acquired using the built-in mechanical classification and biomass estimation
shutter. Due to the time-intensive nature of HSI, cap- The acquisition setting that consistently produced the
turing images of many biological replicates under dif- best classifications across all three species was chosen to
ferent settings on the same day was not feasible. Thus, construct a robust algorithmic model for each configura-
for CONF1, two biological replicates per species were tion. At least two images per species were used to train
imaged at five time points over the course of twenty days. each model to ensure that each species was equally rep-
In CONF2, two biological replicates per species were resented. Raw datacubes and their corresponding classi-
imaged at two timepoints over eight days. For CONF3, fied images were imported using the SPy and Pillow (PIL)
three technical replicates of a single biological replicate v9.0.1 [37] modules. For each configuration, two mod-
for each species were imaged on one day. The total num- els were generated: one trained on the SAM classifica-
ber of images are listed in Table 1. tions and the other trained on the K-Means clusters. The
datacubes and class images were cropped to a region of
Hyperspectral image pre‑processing, classification, 300 × 300 pixels at the root-soil interface, encompassing
and data reduction all three classes. These datasets were reshaped into two-
An overview of the data processing workflow is repre- dimensional dataframes, where each pixel represented
sented in Fig. 2. Reflectance corrections for the white a row, and the selected bands served as columns. Each
and dark reference were carried out automatically in the pixel was assigned to one of the three classes based on
HSI SNAPSCAN software. After this, each image was the respective classified image. All data for each model
classified using a supervised and an unsupervised clas- were consolidated into a single dataframe, and rows
sification method. The supervised Spectral Angle Map- containing unlabeled pixels from the SAM classification
per (SAM) classifications were also carried out in the were excluded.
HSI SNAPSCAN software. Three different regions of RandomForestClassifier (RF) and SVC (Support Vec-
interest (ROI) were manually selected for root, soil, and tor Machine Classification, SVM) from the Scikit-learn
the root-soil interface, with a minimum of 100 pixels per v1.0.2 module [38] were implemented for machine learn-
ROI. These ROIs were used to run the SAM classifica- ing. The dataframes were randomly split into a training
tion with a spectral angle of 10 degrees. All remaining set (80%) and a testing set (20%) using a random state
data processing was carried out in Python v3.2. Unsu- of 0 and either a fixed number of estimators (50) for the
pervised K-Means clustering was carried out using the RF model, or a linear kernel for the SVM model. After
Spectral Python (SPy) v0.23.1 module [32]. After the raw being fitted using the assigned classes, the models were
Faehn et al. Plant Methods (2024) 20:171 Page 6 of 17

Fig. 2 Data processing workflow for hyperspectral image analysis. Image pre-processing and SAM classifications were carried out in the HSI
SNAPSCAN software. All other steps were carried out in Python

used to predict the testing set. Models were evaluated hairs within the organic mucilaginous matrix of the
based on the classification_report, confusion_matrix, rhizosheath, was crucial to clearly differentiate between
and accuracy_score metrics from the Scikit-learn mod- root and soil classes due to its heterogeneity and the spa-
ule. The models developed were used to classify the root, tial resolution limits of the camera.
soil, and interface regions across all rhizobox images in The differences in root architecture between the three
each respective configuration and evaluated based on species, as well as the different configuration (CONF)
the same metrics stated above. The spectral signatures of and acquisition settings chosen, had a clear impact on
each class were used to assess the differences between the classification performance. For CONF1, which gave the
configurations, acquisition parameters, and species. A broadest view of the entire root system, the roots of A.
Partial Least Squares (PLS) regression [39] was used to odoratum and E. vaginatum were well-established and
evaluate the correlations between the predicted classes classified with great accuracy, while the root system of D.
with each species and configuration. flexuosa seemed less vigorous and was not as accurately
For the model predicted images in CONF1, the num- classified (Fig. 3A). In addition, images for all three spe-
ber of pixels in each class were converted to percentages cies taken with larger apertures (f/2.8 and f/4) suffered
using the PIL and webcolors v1.13 modules. The dimen- from overexposure and poor classification by both SAM
sions of the images (115 × 155 mm) were then used to classifications and K-Means clustering, which skewed the
calculate the estimated biomass area for each class. The resulting spectra (Supplementary Fig. 1). After removal
scripts for data analysis are available from GitHub (see of the poor spectral data, only twelve images from an
Availability of Data and Materials). aperture of f/5.6 were used for further data processing.
For CONF2 that was designed with the rhizoboxes
Results positioned closer to the lens, the SAM classifications
Classification challenges of the root, soil, and root‑soil exhibited slight differences in pixels assigned to the spe-
interface in different configurations cific classes with different acquisition settings, while
Three regions of interest (ROI), determined by group- there were no discernible differences with K-Means
ing pixels with similar spectral features, were utilized clustering. With SAM classifications, a time delay inte-
by two classification methods, Spectral Angle Mapper gration (TDI) of 5 classified fewer soil pixels compared
(SAM) classifications and K-Means clustering, to distin- to the standard conditions (TDI of 1), and an aperture
guish root, soil, and root-soil interface regions in images of f/8 classified more soil pixels than an aperture of f/11
of root systems grown in rhizoboxes. The interface class, (Fig. 3B). The difference between a pixel blur of 0 or 2.5
designed to include inorganic soil particles and live root did not lead to any significant changes.
Faehn et al. Plant Methods (2024) 20:171 Page 7 of 17

clustering was only able to distinguish between two

classes at a magnification of 1.6x and 4x for most images
(images for 1.6x not shown due to similarity with a mag-
nification of 4x). This was due to the presence of much
more fine details and specific variation between different
root sections. For the few images where the root class was
detected, the resulting spectral signatures were incoher-
ent (Supplementary Fig. 2). Therefore, these images were
removed from further processing for the K-Means data.
In contrast, SAM classification was able to distinguish all
three spectral classes for all three tested magnifications
(Fig. 3C). The images used for each of the classification
methods and further data processing are listed in Table 1.

Utilizing the second derivative for spectral variable

selection enables effective data reduction for machine
learning
To identify relevant spectral features and decrease the
data size for executing machine learning algorithms, the
root spectra was selected to identify the most informa-
tive bands. When evaluating the optimal window size for
Savitzky-Golay (SG) smoothing using the second deriva-
tive, a window of 21 wavelengths proved effective in fil-
tering out noise, while preserving the integrity of signal
bands. Smaller windows tended to retain artifact signals,
while larger windows exceeding 25 wavelengths may have
over-smoothed genuine sample signals (Supplementary
Fig. 3). This parameter allowed the data to be reduced
from 150 to 16 bands for SAM (Fig. 4A) and 15 bands
for K-Means spectra (Fig. 4B), to facilitate expedited
processing of machine learning algorithms. The selected
wavelengths between the two classification methods
slightly differed in the 545 nm—726 nm range.

Random Forest (RF) models trained on SAM classifications

achieve higher accuracy compared to training on K‑Means
labels
Fig. 3 Classification of root systems of three graminoid species
in three tested configurations. A CONF1 at an aperture of f/5.6
Of the two common machine learning algorithms tested,
with the following acquisition settings: TDI 3, Pixel blur 2.5, Binning Random Forest classifier (RF) and Support Vector
2 × 2. Two biological replicates for each species were employed. B Machine (SVM), the SVM model took more computa-
CONF2 with the acquisition settings of TDI 1, Pixel blur 0, Binning tional time with similar results to the RF model, and thus
1 × 1, unless otherwise indicated. One biological replicate for each the RF model was further used to develop a robust clas-
species was imaged under the different acquisition settings. C
CONF3 with the following acquisition settings: TDI 1, pixel blur 0,
sification model for root systems. For each configuration,
binning 1 × 1. One biological replicate for each species was imaged two RF models were developed: one trained on the SAM
at the different magnifications. In all configurations, the RF model classifications and the other trained on the K-Means
was trained using SAM classifications. A. odor. = A. odoratum, D. classifications, with equal representation of all three
flex. = D. flexuosa, E. vagi. = E. vaginatum graminoid species in the training dataset. One param-
eter within each configuration was selected to train the
For CONF3 that was tailored to capture magnified models. The chosen parameters were an aperture of f/5.6
sections of various regions within each root system, the for CONF1, an aperture of f/11 for CONF2, and a mag-
accuracy of classifications became less discernible at nification of 0.63 x for CONF3. Cropping the datacubes
higher magnifications (Fig. 3C). Specifically, K-Means and associated classified images to a 300 × 300 pixel area
encompassing all three classes, and reducing the datasets
Faehn et al. Plant Methods (2024) 20:171 Page 8 of 17

Fig. 4 Wavelength selection by Savitzky-Golay smoothing. The second derivatives of the root spectra for all images that were processed
in the three configurations of all three graminoid species are shown for (A) Spectral Angle Mapper (SAM) classifications and (B) K-Means clustering.
The wavelengths were selected from the peaks and troughs using the average of each dataset

to the selected wavelengths (Fig. 4), facilitated a more F1-scores was found to vary across different configura-
efficient model training process. Additionally, unlabeled tions and acquisition parameters (Fig. 5).
pixels in the datasets classified using SAM were removed. For CONF1 (Fig. 5A), the root class exhibited the high-
The results from testing the models, indicated that the est variability and lowest accuracy compared to the other
RF models for the SAM classified images had an 88–91% configurations (Fig. 5B and C). For CONF2 (Fig. 5B),
accuracy (macro average of all per-class F1-scores), while images taken at an aperture of f/11 with a pixel blur of
for the K-Means classified images the accuracy was 2.5 along with an aperture of f/8 showed higher accuracy
77–85% (Table 2), which resulted in more pixels being for all three classes than the images taken at an aperture
inaccurately predicted to be interface or soil with the of f/11 with the standard settings. Since the latter were
K-Means RF model (Supplementary Fig. 4). Therefore, the images on which the model was trained, this find-
the SAM trained RF model was considered to be supe- ing was noteworthy. Additionally, the images taken at an
rior to the K-Means RF model and was used to predict all aperture of f/11 with a TDI of 5 had the lowest accuracy.
other images in each configuration. For CONF3 (Fig. 5C), a magnification of 0.63x achieved
the highest accuracy for all three classes, whereas
Model accuracy varies between imaging configurations higher magnifications had lower accuracies. This might
and data acquisition parameters be attributed to the fact that the model was trained on
To evaluate the effects of settings selected during image the dataset it performed on best. However, training the
acquisition, the following parameters were methodically model using data from all three magnfications led to
adjusted in each configuration: apertures in CONF1; inaccuracies (Supplementary Fig. 5) and was therefore
apertures, TDI, and pixel blur settings in CONF2; and not feasible. This suggests that each magnification in
magnification lenses in CONF3. Only the accuracy scores CONF3 may require a distinct, customized classification
for each of the three classes were analyzed since inclu- model for effective data analysis.
sion of unlabelled pixels in the original classified images Generally, the accuracy scores during prediction were
skewed the calculation for the overall accuracy (macro lower than the testing scores obtained during model
average). The prediction accuracy in each class as mean training for all configurations. For instance, the root class
Faehn et al. Plant Methods (2024) 20:171 Page 9 of 17

Table 2 Classification accuracy for Random Forest (RF) models trained on Spectral Angle Mapper (SAM) classifications and K-Means
classifications
CONF Classification Featurea 1. Soil 2. Interface 3. Root Macro avg

CONF1 (aperture: f/5.6) SAM Precision 0.89 0.86 0.94 0.9

Recall 0.97 0.73 0.88 0.86
F1-scoreb 0.93 0.79 0.91 0.88
Support 27307 12871 5608 45786
K-Means Precision 0.85 0.75 0.89 0.83
Recall 0.85 0.79 0.56 0.73
F1-scoreb 0.85 0.77 0.69 0.77
Support 58132 42645 7223 108000
CONF2 (aperture: f/11) SAM Precision 0.99 0.89 0.86 0.91
Recall 0.99 0.9 0.83 0.91
F1-scoreb 0.99 0.89 0.85 0.91
Support 19233 18884 12,074 50,191
K-means Precision 0.91 0.76 0.92 0.86
Recall 0.86 0.85 0.8 0.84
F1-scoreb 0.88 0.81 0.86 0.85
Support 90133 59002 12865 162000
CONF3 (magnification: 0.63x) SAM Precision 0.95 0.83 0.93 0.90
Recall 0.96 0.81 0.92 0.90
F1-scoreb 0.96 0.82 0.92 0.90
Support 74209 30945 30065 135219
K-means Precision 0.86 0.71 0.9 0.82
Recall 0.83 0.77 0.83 0.81
F1-scoreb 0.84 0.74 0.86 0.81
Support 68874 56367 36759 162000
a
The precision, recall, F1-scores, and support (number of actual pixels) for each of the three classes (Soil, Interface, Root), and the macro average (arithmetic mean of
all per-class F1-scores) for each of the testing datasets for the RF models is provided for each configuration (CONF)
b
The F1-score, which is a harmonic mean of the precision and recall values, is used as the main metric of accuracy on a scale of 0 to 1, where the closer it is to 1
represents a precise and accurate model

in CONF2 had a testing accuracy of 85% but a predicted vaginatum had the highest intensity in all three configu-
accuracy of only 69% ± 15% for the same aperture on rations for the root class, while D. flexuosa had the lowest
which the model was trained. Similarly, for CONF3 the intensity, though it’s spectra was similar to A. odoratum
testing accuracy was 92%, but the predicted accuracy was in CONF3. E. vaginatum had the highest intensity for all
86% ± 7% on the same images. Despite these variations, three classes in CONF3. CONF2 produced the most vari-
the RF models outperformed either of the previous clas- ation in spectral signatures within the 550–700 nm range
sification methods for all images (Fig. 3). Due to this, the for all three classes compared to CONF1 and CONF3,
quality of the spectral signatures for each image needed and the spectral signatures between the three species
to be compared to identify the differences between con- were similar, only varying in intensity.
figurations, acquisition parameters, and species. The different parameters within each configuration also
influenced the spectral signatures (Fig. 6D–F). Spectral
Different imaging configurations have the largest impact signature characteristics remained consistent between
on spectral signatures apertures f/8.0 and f/11.0 in CONF2, only varying in
Spectral analysis showed that the different configura- intensity. A TDI of 5 at an aperture of f/11 in CONF2
tions had the largest effect on spectral signatures, while resulted in smoother signatures with lower intensity for
the different graminoid species had differences in inten- the root class. In CONF3, a magnification of 4x had the
sity across all three classes within each configuration highest intensity for all three classes, but a smoother
(Fig. 6A–C). CONF1 and CONF3 produced smoother spectral signature. A Partial Least Squares (PLS) regres-
spectral signatures than those in CONF2, and CONF1 sion confirmed that the three classes correlated most
had the lowest intensity of all three configurations. E. strongly, irrespective of species (Fig. 6G), however, the
Faehn et al. Plant Methods (2024) 20:171 Page 10 of 17

Fig. 5 RF model prediction accuracy. The prediction accuracy for each individual class and the macro average accuracy of the combined
classes for each acquisition parameter tested within each configuration: A CONF1, B CONF2, and C CONF3. The accuracy is represented
by the average ± standard deviation of the F1-scores are as follows: for CONF1, f/2.8 (N = 18), f/4 (N = 18), f/5.6 (N = 30); for CONF2, f/8 (N = 8), f/11
(N = 16), f/11 TDI 5 (N = 3), f/11 pixel blur 2.5 (N = 3); and for CONF3, each magnification (N = 9)

three configurations also correlated within each class the observed period (20 days), the root systems of A.
(Fig. 6H). odoratum had the highest increase in estimated root
biomass over the time course (Fig. 7A), however there
was substaintial variance between the two biologi-
Estimation of root biomass with HSI
cal replicates. The estimated biomass for the interface
Since CONF1 provided a full overview of the root sys- classes showed a greater increase than that of the root
tems, it was used to estimate root biomass alongside class (Fig. 7B).
their spectral signature characterization by HSI. Over
Faehn et al. Plant Methods (2024) 20:171 Page 11 of 17

Fig. 6 Spectral signatures and Partial Least Squares (PLS) regression of all RF model predicted spectra. The average reflectance spectra across all
wavelengths in each class for (A-C) each species and configuration (CONF) and (D-F) each acquisition parameter and configuration. The predicted
spectra for all data were fitted to (G) the combination of each species and class, and (H) the combination of each configuration and class

Discussion and classification is a vital aspect of HSI analysis,

To evaluate the strengths, weaknesses, potentials, and which have proven to be challenging for agricultural
pitfalls of HSI technology in root biological applica- tasks such as detecting disease and pest damage on
tions, we used the imec VNIR SNAPSCAN camera to leaves [41, 42], or identifying weeds in crops fields
image rhizobox-grown root systems of three anatomi- [43]. Segmenting roots from soil presents even greater
cally different graminoid species. This camera, with its challenges due the heterogenous nature of soil and,
high spatial resolution of up to 3650 × 2048 pixels, was depending on the soil type, the potential spectral simi-
chosen for its ability to capture detailed images com- larities between dry soil and living roots or wet soil and
pared to other NIR or MWIR HSI cameras that uti- dead roots [18, 44]. Thus, high resolution hyperspectral
lize linescan technology. Although these cameras may images are necessary to accurately classify soil-grown
offer wider spectral resolutions, their spatial resolu- root systems.
tions are typically more limited [40]. Additionally, the In our study, we focused on distinguishing roots
ability to mount the imec camera on a microscope fur- and soil as the two most obvious ROIs. In addition,
ther enhances its capability to distinguish between the we attempted to classify the root-soil interface/rhizos-
root system and surrounding soil. Image segmentation heath where root exudates change the physical and
Faehn et al. Plant Methods (2024) 20:171 Page 12 of 17

Fig. 7 Estimated surface biomass for the root and interface classes. The estimated biomass for the (A) root, and (B) interface classes in CONF1
over the 20-day imaging period. Biomass was calculated from the area of the image and percentage of pixels in each class. Each point
is the average ± standard error of two biological replicates

physiological characteristics of the soil, which is a comparative analysis of spectral signatures across dif-
region of high biological and biotechnological inter- ferent configurations, acquisition parameters, and
est [45, 46]. However, due to the complex nature of the species.
rhizosheath, this ROI was not well resolved, either spa-
tially or spectrally, and included varying proportions of Imaging configuration and acquisition parameters
root and soil areas across all images. To our knowledge, influence the reflectance spectra
there have been no previous attempts to use HSI for For optimal data capture, selecting the appropriate imag-
detecting the rhizosheath. However, there is evidence ing configuration and acquisition settings tailored to
suggesting that the presence of fungi, mold, algae, and the specific application is crucial. In spectral cameras
differences in soil water potential can be classified to equipped with an integrated thin-film Fabry–Perot fil-
provide insights into the interactions between roots ter, such as the imec SNAPSCAN camera, the position of
and the rhizosphere [17, 18]. Despite its challenges, each filter on the sensor and the aperture size can shift
the interface ROI was kept as it helped to increase the the measured spectra, potentially resulting in a loss of
distinction between the root and soil classes. Utiliz- detail, particularly with larger apertures [47]. This may
ing three classes throughout our analyses enabled a be the reason for the loss of spectral detail in CONF1 in
Faehn et al. Plant Methods (2024) 20:171 Page 13 of 17

comparison to CONF2, however, the smoother signa- the estimated surface-level root biomass from classifica-
tures obtained for CONF1 could also be attributted to tion outcomes and the actual biomass was not explored
the images being binned (2 × 2) and having a Time Delay or confirmed in this study. This point is shown to dem-
Integration (TDI) of 3. Pixel binning can be set from 1 onstrate the method’s potential applicability in future
to 20 to merge pixels in a NxN neighborhood, where the research, particularly in studies focused on the tradeoffs
benefits of binning can increase the signal to noise ratio between root trait plasticity and belowground biomass
(SNR), which is optimal for larger ROIs to reduce individ- allocation as key indicators of plant resilience to environ-
ual pixel noise [48], but will inevitably decrease the spa- mental stresses [51].
tial resolution. TDI sets the step size (1–5) for imaging In CONF2, positioning the sample closer to the lens
each spectral band, where lower values allow informa- allows for a more detailed analysis of specific root system
tion for pixels in the same spectral band to be averaged segments with clear distinctions in spectral signatures.
across multiple frames, thus significantly improving the This proximity enhances resolution and detail, which
SNR [49]. The effect of no TDI (a setting of 5) was seen is essential for linking HSI data with specific root func-
in CONF2 which resulted in smoother spectral signa- tional traits and compositional variations. The differences
tures with lower intensity. In CONF3, using the highest in spectral signatures between the root and soil classes
magnification factor (4x) led to a loss of detail within the were similar to other studies, where root signatures had
spectral signature. This outcome could be expected, con- higher average reflectance and greater variance than the
sidering that the spectral data from other configurations surrounding soil [18]. When combined with deep-learn-
represented averages across larger regions of the root ing models, this configuration is optimal for classifying
system, while at 4x magnification, the spectral signature chemical or nutrient concentrations and detecting related
was derived from just 0.25 ­cm2 sections of the root sys- stresses or resilience in root systems [15]. Although this
tem. The variation in spectral signatures across different study primarily focused on comparing imaging configu-
acquisition settings highlights the necessity of maintain- rations and acquisition settings, the insights gained lay
ing consistent conditions throughout a study to ensure a foundation for future research. These studies could
comparable results. explore the relationships between root traits and envi-
ronmental adaptations, thereby deepening our under-
Different imaging configurations can be utilized for various standing of root strategies and their ecological impacts
biological inquiries during environmental changes [5, 52].
The three configurations used in this study are adaptable Future research should focus on refining the set-up and
for various biological questions, with the choice of acqui- data analysis methods for CONF3 using the stereomi-
sition parameters being contingent on the specific appli- croscope. Although distinguishing the root system from
cation. In CONF1, where the sample is positioned at the the soil was achievable, accurately capturing the intricate
greatest distance from the lens, an overview of the root details of smaller root segments might require a more
system architecture is achievable, which can be utilized extensive classification system. With adjustments to the
to track general changes in composition and biomass imaging set-up and analytical approach, this configura-
allocation over time. However, the spectral signatures tion holds potential for detailed investigations of fine-
derived from the acquisition parameters used here scale root-soil interactions and the role of root exudates
should be viewed with skepticism. The binning, TDI, and in the rhizosphere [17, 52]. Furthermore, adapting this
lens aperture settings may not be optimal for this config- set-up for use with a brightfield microscope could yield
uration and should be further investigated. Additionally, significant insights into root physiological processes, a
biomass estimates from these images should be viewed topic that has not yet been thoroughly explored in the
as approximations since they only capture the visible sur- realm of HSI. The purpose of employing CONF3 in this
face portion and exclude the biomass hidden within the study was to demonstrate its capabilities and to bench-
0.6 cm depth of the rhizoboxes. Since the interface class mark it against broader-scale configurations.
generally consisted of a heterogenous mix of fine roots,
root hairs, and soil, interpreting its biomass in a biologi- Choosing image‑processing techniques depends
cal context here is not feasible. Nevertheless, it highlights on the objective for hyperspectral data analysis
that these parts of the root systems contribute to the When processing hyperspectral images, a variety of
overall root biomass. Given the crucial role these regions analytical techniques are available, and selecting the
play in rhizosheath formation and their complex interac- appropriate steps depends on both the priority of the
tions with the surrounding soil [45, 50], this class holds data outcome and the availability of methods. Since our
potential for further investigation into root-soil inter- goal was to assess both the ability to distinguish root
face dynamics. Furthermore, the correlation between systems from the soil matrix, as well as the influence of
Faehn et al. Plant Methods (2024) 20:171 Page 14 of 17

configuration and acquisition parameters on data qual- wavelengths where there are true sample differences and
ity, a data-processing pipeline was developed based on assist in reducing artifact noise in the spectra. Using
common techniques to reduce bias, enhance reproduc- these wavelengths as the method for variable selection
ibility, and increase efficiency. Utilizing the supervised facilitated a significant reduction in data, from 150 to
Spectral Angle Mapper (SAM) classification from the 15 wavelengths. This enabled the effective application of
SNAPSCAN software, which groups pixels by spectral machine learning to evaluate the accuracy of models in
angle similarity to the reference vector [53], alongside predicting root, soil, and interface regions within images.
the widely used unsupervised classification method, Alternatively, Principal Component Analysis (PCA)
K-Means clustering [54], allowed for a quick assessment could have been a suitable method to reduce the dimen-
of root classification accuracy. In general, both classifi- sionality of the data. However, the optimal wavelengths
cation methods exhibited advantages and limitations. between PCA loadings and second derivative spectra
The supervised SAM classification method captured have shown similarity between different sample sets [58],
very fine details of the root system; however, it fell short therefore variable selection by the second derivative was
in classifying all pixels in the image. On the other hand, an effective choice for data reduction.
unsupervised K-Means clustering was more automated
than SAM, classifying each image into three classes Machine learning‑based hyperspectral image classification
with a single line of code, however the detailed root provides a robust method to study root systems
regions identified with SAM were lost and often classi- Both machine learning algorithms initially tested in this
fied as the interface class. Despite K-Means being one of study, Support Vector Machine (SVM) and Random
the most widely used clustering algorithms, it has been Forest (RF) classification, are commonly used classifi-
found to often perform poorly [55], and this seemed to ers [59, 60]. However, due to faster computational time,
be the case here. Although the K-Means algorithm can RF was the preferred algorithm to generate the classifi-
be enhanced with adaptive initialization methods [55], cation models in this study. Consistent with the initial
and more sophisticated clustering algorithms such as classification results, the SAM method produced higher
Artificial Neural Networks (ANN) might surpass SAM accuracy scores when generating the models. Visually,
in performance [56], these approaches typically demand the RF model demonstrated exceptional accuracy in
greater computational resources. These options were not predicting root and soil regions and outperformed the
explored in this study as they were beyond the scope of initial SAM classifications in all configurations. How-
the research. Ultimately, classification accuracy depends ever, the accuracy scores did not reflect this assessment.
on various factors, necessitating further processing steps During model testing, the accuracy for the root class
for de-noising, data reduction, and image compression to ranged from 85–92%, but it dropped to 35–86% when
extract the relevant information [28]. predicting the original images. Beyond the differences
Data reduction is a necessary step for handling the in acquisition settings, the low F1-scores observed
large data size of hyperspectral images to reduce com- in the predicted images can likely be attributed to the
putational load and time. This can be done through model being trained on selectively cropped regions,
spatial or spectral binning, or various variable selection which contained the most accurately classified sections
approaches [57]. The SNAPSCAN software offers spatial of an image. In contrast, predictions were performed
binning capabilities but given the significance of the spa- on entire images, which displayed variations in classifi-
tial variations between the root system and the soil at the cation quality. This discrepancy was particularly notice-
pixel scale, this approach did not appear to be the best able in CONF3, where the stereomicroscope’s lens
method for enhancing spectral distinctions. Calculating limitations resulted in image vignetting. Furthermore,
the second derivative using Savitzky-Golay (SG) smooth- the deviation in accuracy scores within each parameter
ing is a widely used spectral pre-processing technique suggests variability in accuracy over time and across
that enhances desired spectral features while reducing the three species examined. The low accuracy scores
unwanted noise from the sample or instrument, such might also be explained by the evaluation metric of the
as refractive index scattering and white noise [36]. The RF models’ accuracy, which used the assumption that
primary objective of the image processing was to dis- the original SAM labels were accurate. However, these
tinguish root systems from the surrounding soil, so the labels may not accurately represent the true classifica-
second derivative was applied only to the root spectra. tion of each pixel. The SAM classifications were derived
Since there is no general method for selecting the opti- from subjective choices in the manual selection of ROIs
mal window size for SG-smoothing [36], testing a range that constitute only a small fraction of the total pixels.
of window sizes enabled the identification of consistent The classification of the root-soil interface class was
signals across multiple ranges. These signals likely reflect often confounded by variations in the background soil
Faehn et al. Plant Methods (2024) 20:171 Page 15 of 17

or root systems, rather than the interface itself. Add- Author contributions
CF and KKr conceived and planned the research, selected the investigated
ing more classes to differentiate these regions may have plant species and set the scope for the manuscript. CF set up the experiments,
been useful but can also lead to increased subjective performed data analysis and drafted the first manuscript version. GK and MK
bias. Although numerous other model classifiers and supervised and advised technical aspects of HSI and image data analysis. KKr
and KKa contributed to data interpretation. All authors contributed to final-
image processing methods hold the potential of deliver- izing the manuscript and approved its final version.
ing more precise segmentation of root system regions
[12, 18, 61], their exploration was beyond the scope of Funding
Open access funding provided by UiT The Arctic University of Norway (incl
this particular study. Analyzing the spectral signatures University Hospital of North Norway). This research was funded as part of the
for each class was an effective approach to determine strategic project ABSORB by UiT—The Arctic University of Norway. MK was
how various configurations and acquisition settings supported by Research Council of Finland Flagship Programme, Photonic
Research and Innovation (PREIN), decision 346518.
impacted data quality. Future research should build
upon our findings by investigating other methods tai- Availability of data and materials
lored to the specific goals of the analysis. The scripts for data analysis are available from https://​github.​com/​corin​ef/​
Autom​ated-​root-​class​ifica​tion.

Declarations
Conclusions
In this study, we evaluated various HSI acquisition Ethics approval and consent to participate
parameters and data processing techniques for analyzing All methods were carried out in accordance with institutional, national, and
international guidelines and regulations.
root systems across three different imaging scales using
the VNIR SNAPSCAN HSI camera. Our methodology, Consent for publication
which involved use of supervised Spectral Angle Map- Not applicable.
per (SAM) for initial image classification into roots, soil, Competing interests
and the root-soil interface, followed by selection of vari- The authors declare no competing interests.
ables through the use of second derivative and training
Author details
a Random Forest (RF) model, provided a robust frame- 1
Department of Arctic and Marine Biology, The Arctic University of Norway,
work for image classification. This approach effectively 9037 Tromsø, Norway. 2 Department of Life Technologies, University of Turku,
highlighted the spectral signature differences across the 20014 Turku, Finland. 3 Department of Environmental and Biological Sciences,
University of Eastern Finland, 80130 Joensuu, Finland. 4 Center for Photonics
three configurations and acquisition parameters. The Sciences, University of Eastern Finland, 80110 Joensuu, Finland. 5 Arctic Centre
analysis method could be successfully used for all three for Sustainable Energy, The Arctic University of Norway, 9037 Tromsø, Norway.
graminoid species tested, despite differences in their root
Received: 27 May 2024 Accepted: 30 October 2024
architecture. The scripts developed during this study are
available as an online Python-based tool for semi-auto-
mated HSI analysis, offering a scalable framework that
can be expanded and refined with further techniques.
References
1. Ndour PMS, Heulin T, Achouak W, Laplaze L, Cournac L. The rhizosheath:
Supplementary Information from desert plants adaptation to crop breeding. Plant Soil. 2020;456:1–13.
https://​doi.​org/​10.​1007/​s11104-​020-​04700-3.
The online version contains supplementary material available at https://​doi.​
2. Lynch JP. Roots of the second green revolution. Aust J Bot.
org/​10.​1186/​s13007-​024-​01297-x.
2007;55(5):493–512. https://​doi.​org/​10.​1071/​BT061​18.
3. Feeney DS, Crawford JW, Daniell T, Hallett PD, Nunan N, Ritz K, Rivers
Additional file 1: Supplementary Figure 1. The effect of different aper- M, Young IM. Three-dimensional microorganization of the soil–root–
tures on the selection spectra from SAM classifications and K-Means microbe system. Microb Ecol. 2006;52:151–8. https://​doi.​org/​10.​1007/​
clustering for all data from CONF1. Supplementary Figure 2. The resulting s00248-​006-​9062-8.
spectra from K-Means clustering in CONF3 at magnification factors of 1.6x 4. Nwokolo NL, Enebe MC, Chigor CB, Chigor VN, Dada OA. The contribu-
and 4x. Supplementary Figure 3. Window size comparison for the second tions of biotic lines of defence to improving plant disease suppression in
derivative with Savitzy-Golay smoothed root spectra. Supplementary soils: a review. Rhizosphere. 2021;19: 100372. https://​doi.​org/​10.​1016/j.​
Figure 4. Confusion matrix for random forest models trained on Spectral rhisph.​2021.​100372.
Angle mapper classifications or K-Means classifications. Supplementary 5. Bodner G, Alsalem M, Nakhforoosh A. Root system phenotying of soil-
Figure 5. Model predicted image of A. odoratum at 0.63x magnification grown plants via RGB and hyperspectral imaging. Methods Mol Biol.
when the model was trained on all magnification factors. 2021;2264:245–68. https://​doi.​org/​10.​1007/​978-1-​0716-​1201-9_​17.
6. Schmidt JE, Lowry C, Gaudin ACM. An optimized rhizobox protocol to
visualize root growth and responsiveness to localized nutrients. J Vis Exp.
Acknowledgements 2018;140:58674. https://​doi.​org/​10.​3791/​58674.
We acknowledge the Climate Laboratory Holt at UiT for plant care and growth 7. Nagel KA, Putz A, Gilmer F, Heinz K, Fischbach A, Pfeifer J, Faget M,
condition maintenance, and its leader, Prof. Laura Jaakola for fruitful discus- Blossfeld S, Ernst M, Dimaki C, et al. GROWSCREEN-Rhizo is a novel
sions. Strategic funding by UiT The Arctic University of Norway (ABSORB- phenotyping robot enabling simultaneous measurements of root and
Project) is gratefully acknowledged. shoot growth for plants grown in soil-filled rhizotrons. Funct Plant Biol.
2012;39(11):891–904. https://​doi.​org/​10.​1071/​FP120​23.
Faehn et al. Plant Methods (2024) 20:171 Page 16 of 17

8. Bontpart T, Concha C, Giuffrida MV, Robertson I, Admkie K, Degefu T, obtained during autopsy. Spectrochim Acta Part A Mol Biomol Spectrosc.
Girma N, Tesfaye K, Haileselassie T, Fikre A, et al. Affordable and robust 2022;274: 121092. https://​doi.​org/​10.​1016/j.​saa.​2022.​121092.
phenotyping framework to analyse root system architecture of soil- 26. Kotronias RA, Fielding K, Greenhalgh C, Lee R, Alkhalil M, Marin F, Emfiet-
grown plants. Plant J. 2020;103(6):2330–43. https://​doi.​org/​10.​1111/​tpj.​ zoglou M, Banning AP, Vallance C, Channon KM, et al. Machine learning
14877. assisted reflectance spectral characterisation of coronary thrombi cor-
9. Rellán-Álvarez R, Lobet G, Lindner H, Pradier P-L, Sebastian J, Yee M-C, relates with microvascular injury in patients with ST-segment elevation
Geng Y, Trontin C, LaRue T, Schrager-Lavelle A, et al. GLO-Roots: an imag- acute coronary syndrome. Front Cardiovas Med. 2022;9: 930015. https://​
ing platform enabling multidimensional characterization of soil-grown doi.​org/​10.​3389/​fcvm.​2022.​930015.
root systems. Elife. 2015;4: e07597. https://​doi.​org/​10.​7554/​eLife.​07597. 27. Vandenabeele M, Veys L, Lemmens S, Hadoux X, Gelders G, Masin L,
10. Rahman G, Sohag H, Chowdhury R, Wahid KA, Dinh A, Arcand M, Vail S. Serneels L, Theunis J, Saito T, Saido TC, et al. The ­AppNL-G-F mouse retina
SoilCam: a fully automated minirhizotron using multispectral imaging for is a site for preclinical Alzheimer’s disease diagnosis and research.
root activity monitoring. Sensors. 2020;20(3):787. https://​doi.​org/​10.​3390/​ Acta Neuropathologica Commun. 2021. https://​doi.​org/​10.​1186/​
s2003​0787. s40478-​020-​01102-5.
11. Handy G, Carter I, Mackenzie AR, Esquivel-Muelbert A, Smith AG, Yaffar D, 28. Vidal M, Amigo JM. Pre-processing of hyperspectral images. essential
Childs J, Arnaud M. Variation in forest root image annotation by experts, steps before image analysis. Chemometrics Intell Lab Syst. 2012;117:138–
novices, and AI. Plant Methods. 2024;20(1):154. https://​doi.​org/​10.​1186/​ 48. https://​doi.​org/​10.​1016/j.​chemo​lab.​2012.​05.​009.
s13007-​024-​01279-z. 29. Balsbergpahlsson AM. Growth, radicle and root hair development of
12. Bodner G, Nakhforoosh A, Arnold T, Leitner D. Hyperspectral imaging: a Deschampsia flexuosa (L.) Trin. seedlings in relation to soil acidity. Plant
novel approach for plant root phenotyping. Plant Methods. 2018;14:84. Soil. 1995;175:125–32. https://​doi.​org/​10.​1007/​BF024​13017.
https://​doi.​org/​10.​1186/​s13007-​018-​0352-1. 30. Chen W, Tape KD, Euskirchen ES, Liang S, Matos A, Greenberg J, Fraterrigo
13. Park E, Kim Y-S, Faqeerzada MA, Kim MS, Baek I, Cho B-K. Hyperspectral JM. Impacts of arctic shrubs on root traits and belowground nutri-
reflectance imaging for nondestructive evaluation of root rot in Korean ent cycles across a northern Alaskan climate gradient. Front Plant Sci.
ginseng (Panax ginseng Meyer). Front Plant Sci. 2023. https://​doi.​org/​10.​ 2020;11: 588098. https://​doi.​org/​10.​3389/​fpls.​2020.​588098.
3389/​fpls.​2023.​11090​60. 31. Gould B, McCouch S, Geber M. De novo transcriptome assembly and
14. Shao Y, Ji S, Xuan G, Ren Y, Feng W, Jia H, Wang Q, He S. Detection and identification of gene candidates for rapid evolution of soil Al tolerance
analysis of chili pepper root rot by hyperspectral imaging technology. in Anthoxanthum odoratum at the long-term park grass experiment. PLoS
Agronomy. 2024;14(1):226. https://​doi.​org/​10.​3390/​agron​omy14​010226. ONE. 2015;10(7): e0124424. https://​doi.​org/​10.​1371/​journ​al.​pone.​01244​
15. Zhou X, Zhao C, Sun J, Yao K, Xu M. Detection of lead content in oilseed 24.
rape leaves and roots based on deep transfer learning and hyperspectral 32. Boggs T, March D, McGibbney LJ, Magimel F, Mason G, Banman K, Jouni
imaging technology. Spectrochim Acta Part A Mol Biomol Spectrosc. M, Kumar R, et al. The Gitter Badger spectralpython/spectral: spectral
2023;290: 122288. https://​doi.​org/​10.​1016/j.​saa.​2022.​122288. Python (SPy). 2022. Zenodo. https://​doi.​org/​10.​5281/​zenodo.​71350​91.
16. Xu Z, Hu H, Wang T, Zhao Y, Zhou C, Xu H, Mao X. Identification of growth 33. The Pandas Development Team. Pandas-dev/pandas: Pandas (v203).
years of Kudzu root by hyperspectral imaging combined with spectral– 2023. Zenodo. https://​doi.​org/​10.​5281/​zenodo.​80927​54.
spatial feature tokenization transformer. Comput Electron Agric. 2023;214: 34. Savitzky A, Golay MJ. Smoothing and differentiation of data by simplified
108332. https://​doi.​org/​10.​1016/j.​compag.​2023.​108332. least squares procedures. Anal Chem. 1964;36(8):1627–39. https://​doi.​
17. Nakaji T, Noguchi K, Oguma H. Classification of rhizosphere components org/​10.​1021/​ac602​14a047.
using visible–near infrared spectral images. Plant Soil. 2008;310:245–61. 35. Virtanen P, Gommers R, Oliphant TE, Haberland M, Reddy T, Cournapeau
https://​doi.​org/​10.​1007/​s11104-​007-​9478-z. D, Burovski E, Peterson P, Weckesser W, Bright J, et al. SciPy 1.0: funda-
18. Chang SJ, Chowdhry R, Song Y, Mejia T, Hampton A, Kucharski S, Sazzad mental algorithms for scientific computing in Python. Nat Methods.
TM, Zhang Y, Koppal SJ, Wilson CH, et al. HyperPRI: a dataset of hyper- 2020;17:261–72. https://​doi.​org/​10.​1038/​s41592-​019-​0686-2.
spectral images for underground plant root study. Comput Electron 36. Zimmermann B, Kohler A. Optimizing Savitzky-Golay parameters for
Agric. 2024;225: 109307. https://​doi.​org/​10.​1016/j.​compag.​2024.​109307. improving spectral resolution and quantification in infrared spectroscopy.
19. Song Y, Sapes G, Chang S, Chowdhry R, Mejia T, Hampton A, Kucharski Appl Spectrosc. 2013;67(8):892–902. https://​doi.​org/​10.​1366/​12-​06723.
S, Sazzad TMS, Zhang Y, Tillman BL, et al. Hyperspectral signals in the 37. Clark A. Pillow (PIL Fork) documentation. Readthedocs. 2015. https://​build​
soil: plant–soil hydraulic connection and disequilibrium as mecha- media.​readt​hedocs.​org/​media/​pdf/​pillow/​latest/​pillow.​pdf. Accessed 1
nisms of drought tolerance and rapid recovery. Plant, Cell Environ. Jul 2023.
2024;47(11):4171–87. https://​doi.​org/​10.​1111/​pce.​15011. 38. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O,
20. Großkinsky DK, Svensgaard J, Christensen S, Roitsch T. Plant phenomics Blondel M, Prettenhofer P, Weiss R, Dubourg V, et al. Scikit-learn: machine
and the need for physiological phenotyping across scales to narrow the learning in Python. J Machine Learn Res. 2011;12:2825.
genotype-to-phenotype knowledge gap. J Exp Bot. 2015;66(18):5429–40. 39. Kleinbaum DG, Kupper LL, Muller KE, Nizam A. Applied regression analysis
https://​doi.​org/​10.​1093/​jxb/​erv345. and multivariable methods. 3rd ed. Pacific Grove, Ca: Brooks/Cole Publish-
21. Jammer A, Akhtar SS, Amby DB, Pandey C, Mekureyaw MF, Bak F, Roth PM, ing Company; 1998.
Roitsch T. Enzyme activity profiling for physiological phenotyping within 40. Manley M. Near-infrared spectroscopy and hyperspectral imag-
functional phenomics: plant growth and stress responses. J Exp Bot. ing: non-destructive analysis of biological materials. Chem Soc Rev.
2022;73(15):5170–98. https://​doi.​org/​10.​1093/​jxb/​erac2​15. 2014;43(24):8200–14. https://​doi.​org/​10.​1039/​C4CS0​0062E.
22. Liu K-H, Yang M-H, Huang S-T, Lin C. Plant species classification based on 41. Krüger M, Zemanek T, Wuttke D, Dinkel M, Serfling A, Böckmann E.
hyperspectral imaging via a lightweight convolutional neural network Hyperspectral imaging for pest symptom detection in bell pepper. Plant
model. Front Plant Sci. 2022;13: 855660. https://​doi.​org/​10.​3389/​fpls.​2022.​ Methods. 2024;20:156. https://​doi.​org/​10.​1186/​s13007-​024-​01273-5.
855660. 42. Singh V, Misra AK. Detection of plant leaf diseases using image segmen-
23. Garillos-Manliguez CA, Chiang JY. Multimodal deep learning and visible- tation and soft computing techniques. Inform Proc Agricu. 2017;4(1):41–
light and hyperspectral imaging for fruit maturity estimation. Sensors. 9. https://​doi.​org/​10.​1016/j.​inpa.​2016.​10.​005.
2021;21(4):1288. https://​doi.​org/​10.​3390/​s2104​1288. 43. Liu T, Zhao Y, Wang H, Wu W, Yang T, Zhang W, Zhu S, Sun C, Yao Z.
24. Amziane A, Losson O, Mathon B, Dumenil A, Macaire L. Reflectance Harnessing UAVs and deep learning for accurate grass weed detection in
estimation from multispectral linescan acquisitions under varying wheat fields: a study on biomass and yield implications. Plant Methods.
illumination—application to outdoor weed identification. Sensors. 2024;20:144. https://​doi.​org/​10.​1186/​s13007-​024-​01272-6.
2021;21(11):3601. https://​doi.​org/​10.​3390/​s2111​3601. 44. Zhang Y, Sazzad TM, Song Y, Chang SJ, Chowdhry R, Mejia T, Hampton A,
25. Brunner A, Schmidt VM, Zelger B, Woess C, Arora R, Zelger P, Huck CW, Kucharski S, Gerber S, Tillman B, et al. 2024 Cost-efficient active illumina-
Pallua J. Visible and near-infrared hyperspectral imaging (HSI) can reliably tion camera for hyperspectral reconstruction. arXiv:​2406.​19560​v1. https://​
quantify CD3 and CD45 positive inflammatory cells in myocarditis: pilot doi.​org/​10.​48550/​arXiv.​2406.​19560.
study on formalin-fixed paraffin-embedded specimens from myocard 45. Aslam MM, Karanja JK, Dodd IC, Waseem M, Weifeng X. Rhizosheath: an
adaptive root trait to improve plant tolerance to phosphorus and water
Faehn et al. Plant Methods (2024) 20:171 Page 17 of 17

deficits? Plant, Cell Environ. 2022;45(10):2861–74. https://​doi.​org/​10.​1111/​

pce.​14395.
46. Galloway AF, Knox P, Krause K. Sticky mucilages and exudates of plants:
putative microenvironmental design elements with biotechnological
value. New Phytol. 2020;225(4):1461–9. https://​doi.​org/​10.​1111/​nph.​
16144.
47. Goossens T, Geelen B, Pichette J, Lambrechts A, Van Hoof C. Finite aper-
ture correction for spectral cameras with integrated thin-film Fabry-Perot
filters. Appl Opt. 2018;57(26):7539–49. https://​doi.​org/​10.​1364/​AO.​57.​
007539.
48. Riccioli C, Pérez-Marín D, Garrido-Varo A. Optimizing spatial data reduc-
tion in hyperspectral imaging for the prediction of quality parameters in
intact oranges. Postharvest Biol Technol. 2021;176: 111504. https://​doi.​
org/​10.​1016/j.​posth​arvbio.​2021.​111504.
49. Fan X, Liu C, Liu S, Xie Y, Zheng L, Wang T, Feng Q. The instrument design
of lightweight and large field of view high-resolution hyperspectral
camera. Sensors. 2021;21(7):2276. https://​doi.​org/​10.​3390/​s2107​2276.
50. Galloway AF, Akhtar J, Marcus SE, Fletcher N, Field K, Knox P. Cereal root
exudates contain highly structurally complex polysaccharides with soil-
binding properties. Plant J. 2020;103(5):1666–78. https://​doi.​org/​10.​1111/​
tpj.​14852.
51. Chandregowda MH, Tjoelker MG, Pendall E, Zhang H, Churchill AC,
Power SA. Belowground carbon allocation, root trait plasticity, and
productivity during drought and warming in a pasture grass. J Exp Bot.
2023;74(6):2127–45. https://​doi.​org/​10.​1093/​jxb/​erad0​21.
52. Williams A, Langridge H, Straathof AL, Muhamadali H, Hollywood KA,
Goodacre R, de Vries FT. Root functional traits explain root exudation rate
and composition across a range of grassland species. J Ecol. 2022;110:21–
33. https://​doi.​org/​10.​1111/​1365-​2745.​13630.
53. Kruse FA, Lefkoff AB, Boardman JW, Heidebrecht KB, Shapiro AT, Barloon
PJ, Goetz AFH. The spectral image processing system (SIPS)—interactive
visualization and analysis of imaging spectrometer data. Remote Sens
Environ. 1993;44(2–3):145–63. https://​doi.​org/​10.​1016/​0034-​4257(93)​
90013-N.
54. Sinaga KP, Yang MS. Unsupervised K-Means clustering algorithm. IEEE
Access. 2020;8:80716–27. https://​doi.​org/​10.​1109/​ACCESS.​2020.​29887​96.
55. Celebi ME, Kingravi HA, Vela PA. A comparative study of efficient initializa-
tion methods for the K-Means clustering algorithm. Expert Syst Appl.
2013;40:200–10. https://​doi.​org/​10.​1016/j.​eswa.​2012.​07.​021.
56. Petropoulos GP, Vadrevu KP, Xanthopoulos G, Karantounias G, Scholze
M. A comparison of spectral angle mapper and artificial neural network
classifiers combined with landsat TM imagery analysis for obtaining burnt
area mapping. Sensors. 2010;10(3):1967–85. https://​doi.​org/​10.​3390/​
s1003​01967.
57. De Juan A, Piqueras S, Maeder M, Hancewicz T, Duponchel L, Tauler R.
Chemometric tools for image analysis. In: Salzer R, Siesler HW, editors.
Infrared and Raman Spectroscopic Imaging. 2nd ed. Weinheim: Wiley;
2014. p. 57–110.
58. Kong W, Zhang C, Cao F, Liu F, Luo S, Tang Y, He Y. Detection of sclerotinia
stem rot on oilseed rape (Brassica napus L.) leaves using hyperspectral
imaging. Sensors. 2018;18(6):1764. https://​doi.​org/​10.​3390/​s1806​1764.
59. Sabat-Tomala A, Raczko E, Zagajewski B. Comparison of support vector
machine and random forest algorithms for invasive and expansive
species classification using airborne hyperspectral data. Remote Sens.
2020;12(3):516. https://​doi.​org/​10.​3390/​rs120​30516.
60. Saha D, Manickavasagan A. Machine learning techniques for analysis of
hyperspectral images to determine quality of food products: a review.
Curr Res Food Sci. 2021;4:28–44. https://​doi.​org/​10.​1016/j.​crfs.​2021.​01.​
002.
61. Smith AG, Petersen J, Raghavendra S, Rasmussen CR. Segmentation of
roots in soil with U-Net. Plant Methods. 2020;16:13. https://​doi.​org/​10.​
1186/​s13007-​020-​0563-0.

Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in pub-
lished maps and institutional affiliations.