NIRPlant: A Spectral Plant-Analysis Ecosystem

• NIRPlant is a spectrally organized plant-analysis ecosystem that uses near-infrared, SWIR, and hyperspectral sensing to capture key biochemical and physiological traits in plants.
• It combines portable NIRS, tunable hyperspectral imaging, and specialized optics like SRN metalens to support high-throughput phenotyping, species identification, and dynamic health monitoring.
• The approach emphasizes model interpretability and workflow integration, leveraging neural networks and chemometric pipelines for both crop and soil fertility analysis.

NIRPlant is best understood as an Editor’s term for a family of plant-analysis approaches organized around near-infrared, short-wave infrared, and adjacent spectrally targeted optical measurements for phenotyping, monitoring, diagnostics, and agronomic decision support. In the literature summarized here, the term does not denote a single standardized algorithm. In one representative usage, it functions as shorthand for an interpretable near-infrared / hyperspectral plant-analysis framework trained on the UPWINS spectral library for vegetation phenotyping and species identification (Basener et al., 2024). More broadly, the same technical space includes portable NIRS phenotyping workflows for breeding, tunable hyperspectral NIR reflectance imaging for leaf water dynamics, chlorophyll-sensitive metalens imaging, SWIR-assisted LIBS bioimaging, RGB+NIR plant localization, and NIR chemometric pipelines for soil carbon and nitrogen estimation (Rife et al., 2021, Stegemann et al., 2024, Khalilian et al., 20 Apr 2025, Kraemer et al., 2017, Kieling et al., 1 Jul 2026). Across these variants, the unifying principle is the exploitation of wavelength-dependent optical signatures associated with pigments, water, proteins, structure, and elemental composition.

1. Conceptual scope and domain boundaries

NIRPlant spans several adjacent but distinct sensing regimes. One branch is portable NIRS for breeding-scale phenotyping, where handheld spectrometers are integrated into organized field and laboratory workflows for high-throughput quality-trait capture (Rife et al., 2021). A second branch is VNIR-SWIR spectral learning, in which reflectance signatures from 400 nm to 2500 nm are used for vegetation phenotyping and species identification, with model interpretability derived from wavelength-wise analysis of learned weights (Basener et al., 2024). A third branch is hyperspectral NIR imaging, exemplified by spectral-phasor hardware operating from 900 nm to 1600 nm for label-free, in vivo monitoring of plant health dynamics (Stegemann et al., 2024).

The domain also includes technologies that sit at the boundary of classical NIR sensing but remain plant-relevant because they target closely related optical mechanisms. Chlorophyll-sensitive metalens systems operate at 685 nm or 660 nm, aligning with chlorophyll absorption peaks for plant health evaluation and polarization-resolved tissue assessment (Khalilian et al., 20 Apr 2025, Khalilian et al., 19 Aug 2025). SWIR excitation at 2090 nm has been used in LIBS bioimaging of plant tissue to improve plasma formation and nutrient mapping in plants grown in lunar regolith simulant (Vozár et al., 30 Apr 2026). This suggests that, in practice, NIRPlant is best treated as a spectrally organized plant-analysis ecosystem rather than a narrow band-limited technique.

A further boundary condition is that NIRPlant is not exhausted by plant-canopy sensing alone. The same literature extends NIR-based inference to soil fertility proxies through portable spectroscopy of Oxisols and Inceptisols, where carbon and nitrogen are estimated from 900–1700 nm spectra via chemometric and machine-learning pipelines (Kieling et al., 1 Jul 2026). A plausible implication is that NIRPlant naturally couples plant and soil analytics when both are measured through field-portable spectral systems.

2. Spectral and biophysical basis

The scientific basis of NIRPlant lies in the fact that plant optical signals encode chemically and physiologically meaningful variation. In the UPWINS-based interpretable neural-network study, spectra with 2152 bands across 400–2500 nm were used to show that learned hidden units emphasize specific spectral regions rather than arbitrary correlations. After removal of inactive neurons, the analysis identified 23 active hidden units, with strong activity in 350–750 nm and additional activity in 1400–1500 nm, 1900–2000 nm, and isolated features near 1700 nm, 2300 nm, and 2500 nm (Basener et al., 2024). The same work interprets separability between Panicum virgatum and Panicum amarum mainly through variation in 350–750 nm likely linked to chlorophyll-a and chlorophyll-b composition, while features near 1850 nm and 2500 nm in Iva frutescens are discussed as consistent with water-related absorption behavior.

NIRPlant should therefore not be reduced to chlorophyll sensing alone. In aboveground crop nitrogen retrieval, a hybrid PROSPECT-PRO + 4SAIL workflow coupled into PROSAIL-PRO showed that the most informative bands for nitrogen retrieval were concentrated in the SWIR, with a best-performing reduced set containing ten bands mainly around 786 nm, 1556 nm, 1568 nm, 1579 nm, 1623 nm, 1656 nm, 1667 nm, 1762 nm, 2124 nm, and 2234 nm (Berger et al., 2020). The same study argues explicitly that nitrogen retrieval should not rely only on chlorophyll proxies because much of plant nitrogen is bound in proteins.

Water-sensitive structure is another major axis. HyperNIR identifies leaf contrast below about 1350 nm with strong structural scattering and around 1450 nm with strong water absorption, enabling phasor-based discrimination of hydration dynamics from only three sequential images (Stegemann et al., 2024). Chlorophyll-sensitive imaging remains important, but in a complementary role: a silicon-rich nitride concentric-ring metalens at 685 nm and a dual-polarization SRN metalens array at 660 nm both leverage strong chlorophyll absorption so that healthier leaves appear darker and chlorophyll-depleted tissue exhibits weaker absorption and higher transmitted intensity (Khalilian et al., 20 Apr 2025, Khalilian et al., 19 Aug 2025).

Taken together, these results establish a multi-parameter spectral logic: visible bands often dominate pigment contrast; NIR and SWIR bands encode water status, structure, and protein-related chemistry; and model performance depends on preserving wavelength-specific information rather than collapsing spectra into single scalar indices.

3. Instrumentation and acquisition architectures

The instrumentation grouped under NIRPlant is heterogeneous, ranging from handheld spectrometers and tunable cameras to metasurface optics and laser-plasma systems.

Modality	Representative system	Reported details
Portable breeding NIRS	Prospector + LinkSquare	Open-source Android app; experiments–samples–scans hierarchy; CSV export compatible with BreedBase
Hyperspectral NIR reflectance imaging	HyperNIR	Three images; LCVR + two linear polarizers; 900–1600 nm; 0.2 hyperspectral cubes/s
Chlorophyll-sensitive flat optics	SRN metalens	685 nm; diameter $40~\mu\text{m}$; measured NA 0.5; focusing efficiency 36%
Polarization-resolved chlorophyll imaging	SRN metalens array	660 nm; simultaneous XLP and YLP acquisition; measured NA 0.54
SWIR-assisted plant bioimaging	LIBS at 2090 nm	Higher SNR, total emissivity, and Mg II / Mg I ratio than 1064 nm
Portable soil NIR	MyNIR	Reflectance mode; 900–1700 nm; 150 scans averaged

In breeding-oriented NIRS, the critical issue is not only the spectrometer but the workflow. Prospector was designed as an open-source Android application for LinkSquare portable NIR spectrometers because manufacturer apps lacked breeding-scale sample organization, barcode-based input, and suitable metadata handling. Its data model is built around experiments, samples, and scans; it stores scan data in an internal SQLite database; it records timestamps and operator identity; and it exports CSV files containing experiment name, scan ID, scan date, device type, device ID, operator, light source, scan notes, and spectral columns in a format adopted by BreedBase (Rife et al., 2021). The ability to set the number of frames per scan from zero to eight is treated as a stability–throughput tradeoff.

HyperNIR represents a different architectural philosophy. Instead of full hyperspectral scanning, it uses a liquid crystal variable retarder between two linear polarizers to synthesize sine-like, cosine-like, and no-filter transmission states, so spectral phasor coordinates are derived from three images. The LCVR can be tuned over 900–1600 nm in windows from 50 nm to 700 nm, and the system preserves full camera resolution while operating at 0.2 hyperspectral cubes per second, limited by the switching rate of the LCVR (Stegemann et al., 2024).

At the compact-optics end, plant health sensing has been demonstrated with CMOS-compatible silicon-rich nitride metasurfaces. A concentric-ring SRN metalens operating at 685 nm achieved measured NA 0.5, focal length $32.5~\mu\text{m}$, measured focusing efficiency 36%, and measured focal spot FWHM $1.42\lambda$, enabling chlorophyll-absorption-based leaf imaging (Khalilian et al., 20 Apr 2025). A later SRN metalens array at 660 nm used orthogonally sensitive sub-apertures for simultaneous X- and Y-linearly polarized transmission imaging, with measured NA 0.54, measured focal length 46.7 $\mu\mathrm{m}$, focal spot FWHM about $1.06\lambda$, and focusing efficiencies of 65.9% for YLP and 55.8% for XLP (Khalilian et al., 19 Aug 2025).

SWIR-assisted LIBS extends NIRPlant into elemental bioimaging. In leaf tissue from broccoli and lettuce grown in lunar regolith simulant and control substrates, 2090 nm excitation produced higher SNR than 1064 nm for Mg I and Ca II, higher total emissivity, and a higher Mg II / Mg I ratio, indicating a hotter and more efficiently ionised plasma (Vozár et al., 30 Apr 2026). Portable soil spectroscopy uses a still simpler setup: the MyNIR device acquired 900–1700 nm reflectance spectra with 150 scans averaged per measurement for carbon and nitrogen modeling in Oxisols and Inceptisols (Kieling et al., 1 Jul 2026).

4. Computational models, inference, and interpretability

A major theme in NIRPlant research is that spectral learning need not be treated as intrinsically opaque. The UPWINS study used a TensorFlow neural network with one hidden layer of 128 ReLU neurons and an 18-neuron softmax output layer, trained with sparse categorical crossentropy for 2000 epochs and batch size 32 on a random 80/20 train-test split (Basener et al., 2024). Although the neural network achieved about 0.87 test accuracy and LDA performed best at about 0.91 accuracy, the emphasis of the paper is interpretability rather than only peak accuracy. The first-layer weights $W^1$ were visualized directly, inactive neurons were removed, and class-specific combinations of first- and second-layer weights were presented as Spectral Activation Plots. The authors argue from these visualizations that the network learns chemical and physiological trait indicators.

This is a notable departure from the common assumption that neural networks on spectral data are irreducibly black boxes. The same paper explicitly states that neural networks can be more explainable than many traditional methods when learned parameters are inspected carefully (Basener et al., 2024). A closely related misconception arises in remote sensing: RGB-pretrained models are often transferred to NIR without addressing domain shift. In response, a ViT + LoRA framework adapted pretrained Vision Transformers to NIR segmentation by freezing the original backbone and training low-rank updates only in the query and value projection layers. The reported effect was roughly a 97% reduction in trainable parameters, with the best NIR performance obtained by LoRA ViT-L/16 + DeepLabV3: IoU 0.884 and F1 0.911 on RIT-18 tree segmentation, and IoU 0.900 and F1 0.912 on DSTL crop segmentation (Ulku et al., 2024). This supports the narrower claim that parameter-efficient adaptation can improve NIR-domain transfer when NIR labels are scarce.

NIRPlant also includes image-based localization. In agricultural robotics, RGB + NIR input was used in a fully convolutional network that regressed a plant location likelihood map for stem emerging point localization. The target map was defined from a Gaussian of the Euclidean distance to the nearest ground-truth SEP, and detections were robustly reproduced with centimeter accuracy in BoniRob experiments (Kraemer et al., 2017). Over a 4-day gap, the landmark comparison reported recall 96.2%, precision 87.5%, and mean error 18.41 mm; over 28 days, recall 88.3%, precision 92.8%, and mean error 20.8 mm (Kraemer et al., 2017). This shows that NIRPlant computation is not limited to spectral regression or classification, but also includes multimodal geometric localization in field environments.

5. Phenotyping, diagnostics, and agronomic use cases

The most immediate NIRPlant application is high-throughput phenotyping in breeding. Portable NIRS is attractive because quality traits are expensive, labor-intensive, and often destructive to measure, so they are frequently deferred to late breeding stages. Prospector addresses the software bottleneck in this workflow by allowing breeders to create experiments, assign samples by typed names or barcodes, collect one or more scans per sample in the field or laboratory, and export organized CSV files for downstream spectral analysis or BreedBase upload (Rife et al., 2021). The breeding significance is operational rather than purely algorithmic: low-cost spectrometers become usable at breeding scale only when metadata, provenance, sample organization, and export formats are handled correctly.

A second use case is vegetation phenotyping and species identification from spectral libraries. The UPWINS framework used 902 spectra measured with an ASD4 field spectrometer across 16 vegetation species and 2 soil types, with rich metadata spanning species, health, growth stage, annual variation, and environmental conditions (Basener et al., 2024). The reported species-identification accuracy is around 90%, but the broader contribution is mechanistic: the network’s learned spectral features are used to expose trait-based spectral separability rather than only assign labels.

A third use case is dynamic plant-health monitoring. HyperNIR demonstrated in vivo water uptake tracking in Capsicum annuum leaves that had been not watered for 3–5 days and were then rewatered with about 40 mL. During the first 15 minutes after watering, the image changed very little; after about 25 minutes, water transport in the capillaries became visible; by about 30–60 minutes, surrounding tissue also showed phasor changes; and after about 1 hour, the rate of change slowed and the image became more uniform (Stegemann et al., 2024). Because these changes were not observed in control experiments without water exposure, the result functions as a proof-of-principle for non-destructive monitoring of rehydration dynamics.

Chlorophyll-sensitive compact optics provide another diagnostic channel. In the 685 nm SRN metalens study, healthy leaves appeared darker because they contained more chlorophyll and therefore absorbed more 685 nm light, while weaker absorption corresponded to reduced chlorophyll levels associated with stress or senescence (Khalilian et al., 20 Apr 2025). The 660 nm polarization-resolved SRN metalens array extended this logic by showing that fusion and difference maps reveal vein boundaries, oriented mesophyll or cell-wall textures, structural anisotropy, and pigment variation across healthy, moderately stressed, early pigment-deficient, and severely chlorophyll-depleted Prunus cerasifera leaves (Khalilian et al., 19 Aug 2025).

NIRPlant can also include elemental monitoring. LIBS bioimaging of broccoli and salad leaves grown for 1 month in lunar regolith simulant or control substrates showed higher Mg and Ca accumulation in the lunar-regolith-grown plants, with tissue-specific localization in veins and midrib for broccoli and more diffuse distribution across salad leaves (Vozár et al., 30 Apr 2026). The plant-monitoring relevance is that a longer SWIR excitation wavelength improved analytical performance while also revealing nutrient uptake from the growth substrate.

Finally, the same methodological family extends belowground into soil fertility estimation. Portable MyNIR spectroscopy combined with Savitzky-Golay preprocessing, NIPALS + Huber-loss outlier removal, and regressors including PLS, SVR, Ridge, Random Forest, and stacking ensembles was used to estimate carbon and nitrogen in Oxisols and Inceptisols (Kieling et al., 1 Jul 2026). Reported validation performance included, for Oxisol carbon, R² = 0.91, RMSE = 0.36, MAE = 0.31, RPD = 3.39 for the best SVR-based pipeline, and for Oxisol nitrogen, R² = 0.89, RMSE = 0.03, MAE = 0.03, RPD = 3.01 (Kieling et al., 1 Jul 2026). These results place NIRPlant within a wider agronomic workflow linking canopy, tissue, and soil measurements.

6. Integration challenges, misconceptions, and research directions

Several recurring limitations define the present state of NIRPlant. The first is that cheap hardware alone is insufficient. The Prospector paper is explicit that affordable handheld spectrometers already existed, but the missing layer was breeding-centric software: organized high-throughput collection, barcode support, metadata handling, and direct export to downstream infrastructure (Rife et al., 2021). A related systems goal is tighter integration with BrAPI-enabled databases and the longer-term formation of a “digital breeding ecosystem,” in which field data capture, analysis, and application are tightly linked (Rife et al., 2021).

A second limitation is dataset scale and transferability. The UPWINS paper argues that current hyperspectral deep learning is limited by small labeled datasets, even while showing that interpretable models can already extract meaningful features from 902 spectra (Basener et al., 2024). The LoRA-ViT paper likewise frames NIR analysis as a domain-shift problem caused by reliance on RGB pretraining and scarcity of annotated NIR data (Ulku et al., 2024). HyperNIR adds a hardware-specific constraint: frame rate is limited by LCVR switching, transmission is not a mathematically perfect sine or cosine in wavelength space, and calibration is required to correct LCVR inhomogeneity (Stegemann et al., 2024).

A third limitation is validation realism. In soil spectroscopy, validation strategy changed the apparent quality of the model; the authors note that very small Kennard-Stone test sets can look artificially good, and they argue that around 30% test was more statistically representative for Oxisol in their study (Kieling et al., 1 Jul 2026). This cautions against treating favorable calibration or overly convenient holdouts as sufficient evidence of field readiness.

Several misconceptions are directly contradicted by the literature. One is that NIRPlant is synonymous with a single vegetation index or with chlorophyll alone; in fact, the cited work spans chlorophyll-sensitive visible bands, NIR structure and water signals, SWIR protein absorption, and elemental bioimaging (Berger et al., 2020, Stegemann et al., 2024). Another is that neural networks necessarily sacrifice mechanistic insight; the UPWINS analysis shows the opposite when the learned weights are treated as wavelength-resolved indicators (Basener et al., 2024).

Future directions are already sketched in the sources. The UPWINS project is intended to expand with additional field spectrometer measurements and eventually hyperspectral imagery from UAVs, aircraft, satellites, and other sensors, along with LiDAR and SAR (Basener et al., 2024). Prospector is envisioned as a frontend that can support new spectrometer models and more direct BrAPI integration (Rife et al., 2021). A plausible implication is convergence between NIRPlant spectral sensing and geometry-centric plant reconstruction. Field NeRF reconstruction already achieves a 74.65% F1 score in outdoor corn scenes against LiDAR ground truth, and PlantSegNeRF provides few-shot, cross-dataset 3D instance point clouds from multi-view RGB image sequences with strong gains in semantic and instance segmentation metrics (Arshad et al., 2024, Yang et al., 1 Jul 2025). Those implementations are not NIR-first, but they indicate the likely direction of a more complete phenotyping stack in which spectral diagnostics and organ-aware 3D geometry are integrated rather than treated separately.

An adjacent systems extension concerns spectral management rather than sensing. A silicon nitride grating-based planar spectral splitting concentrator was designed to harvest 700–1400 nm NIR light while transmitting visible light, with overall NIR guiding efficiency of about 18.41% and electrical conversion efficiency of about 11.27% (E et al., 2020). This does not constitute NIRPlant sensing in the narrow sense, but it suggests that plant-facing NIR photonics may also include control of which spectral bands are reserved for energy harvesting versus illumination.

In aggregate, NIRPlant denotes an emerging technical field defined less by one algorithm than by a common operational premise: plant-relevant physiology, chemistry, structure, and environment can be measured or inferred by treating the near-infrared and adjacent optical spectrum as an information-rich substrate, then coupling acquisition hardware to workflow-aware software and model-based interpretation.