The motivation is rooted in the sheer scale of protected tomato production. According to the Food and Agriculture Organization of the United Nations, annual tomato production in China rose steadily from 2010 to 2024, reaching roughly 61.65 million tons, with yields of about 56,700 kilograms per hectare. The seedling stage is the foundation of that production chain, because leaf physiology reveals nutrient supply, photosynthetic capacity, and water stress long before problems become visible in the fruit. Three indicators matter most: SPAD, a proxy for chlorophyll and photosynthetic potential; nitrogen, which drives chlorophyll, protein, and enzyme formation; and moisture, which governs cell turgor, stomatal regulation, and transport within the plant.

What makes the new work distinctive is that it refuses to treat navigation and sensing as separate problems. Most previous spectral studies relied on fixed platforms, handheld sampling, or offline image acquisition, while most greenhouse robot studies focused purely on positioning and obstacle avoidance. The team, led by Huili Zhang and Yuliang Yun of Qingdao Agricultural University, built a unified workflow in which a mobile platform maps its environment, plans a route, positions itself over a leaf, verifies that the camera is actually looking at healthy leaf tissue, and only then stores a spectrum tied to a timestamped navigation task.

The hardware is deliberately compact. A Panda four-wheel differential-drive chassis measuring 0.47 by 0.35 meters carries a 68,000 mAh battery, an Orange Pi 5 Ultra single-board computer built around a Rockchip RK3588 octa-core processor with a 6-TOPS neural processing unit, a CM020D multispectral camera covering 400 to 950 nanometers in 27 bands, and an RPLIDAR C1 laser scanner with a 0.05 to 12 meter range and ±30 millimeter accuracy. Mapping, localization, path planning, and chassis control are orchestrated through a graphical interface on top of the Robot Operating System, using standard SLAM, adaptive Monte Carlo localization, A* global planning, and dynamic window approach local planning.

One of the most instructive engineering lessons came from the robot’s own body. Four vertical pillars supporting the camera sat close to the LiDAR’s scanning plane, so the laser kept hitting the robot itself. Because SLAM assumes all laser echoes come from the static environment, these self-reflections were repeatedly projected onto the map as phantom obstacles, contaminating the occupancy grid and eventually choking path planning in the narrow aisles. The team’s fix was geometrically simple but effective: they calculated the radial distance of the pillars from the LiDAR origin as roughly 0.178 meters, added a 0.04 meter safety margin, and discarded every echo below a 0.22 meter threshold. After filtering, scattered noise points vanished, aisle boundaries sharpened, and navigation stabilized.

Navigation trials across five experiments showed the platform reaching its targets with a mean success rate of 95.5 percent and an average point-to-point error of just 27.6 millimeters. The team also worked out the minimum aisle width the robot needs to rotate in place — about 0.586 meters before safety margins — a critical figure in greenhouses where seedling benches leave little room to maneuver. Together, these results demonstrated that a small, inexpensive platform could move reliably enough to serve as a stable base for precision spectral acquisition.

The sensing side faced its own subtlety: how do you guarantee the camera is measuring leaf and not background, leaf edges, or shadows? The answer is a four-box green consistency criterion. The camera preview is divided into four fixed subregions at the corners of a central reference area, and the Normalized Difference Vegetation Index — computed from 660 nanometer red and 840 nanometer near-infrared reflectance — is evaluated in each box every 20 milliseconds. A dual-threshold hysteresis scheme marks a box green above an NDVI of 0.28 and red below 0.20, with a low-signal protection rule. Only when all four boxes are simultaneously green does the system reconstruct the full 27-band spectrum, average the four subregions, and save the sample. This simple gate keeps contaminated spectra out of the dataset before they can do any harm.

To convert spectra into physiology, the researchers assembled 200 leaf samples over three days of contrasting weather — cloudy, sunny, and hazy — at a commercial seedling greenhouse in Qingdao, pairing each spectrum with reference readings from a handheld LYS-4N plant nutrition meter. Weather mattered: hazy-day leaves showed systematically higher SPAD, nitrogen, and moisture values than sunny or cloudy ones, and models trained on the smaller cloudy and hazy subsets performed worse. The team therefore built their main models on the 100 sunny-condition samples using a fixed chain: Savitzky–Golay smoothing to suppress noise, standard normal variate transformation to remove scattering and intensity differences between leaves, linear detrending to flatten baseline drift, and competitive adaptive reweighted sampling to distill 27 bands down to the 12 most informative wavelengths.

The payoff was substantial. Preprocessing alone lifted prediction-set R-squared values from 0.35 to 0.81 for SPAD, 0.54 to 0.84 for nitrogen, and 0.57 to 0.84 for moisture. Wavelength selection then cut prediction errors by 16 to 20 percent relative to full-band models while halving the input dimension — and, notably, improved generalization even though training-set fit slightly decreased, a classic signature of reduced overfitting. XGBoost regression outperformed partial least squares, support vector regression, and random forest across all three targets, and a demanding cross-validation scheme that held out entire seedling benches confirmed the models stayed stable on plants they had never seen, with mean R-squared values of 0.81, 0.83, and 0.84.

Finally, the whole system was validated in the real greenhouse. Across five autonomous navigation tasks on two seedling benches, the four-box method triggered 13 to 17 times per bench pass, and the platform’s batch predictions tracked the handheld meter closely: average absolute differences were about 0.55 SPAD units, 0.21 percent for nitrogen, and 0.84 percent for moisture, with biases near zero and no systematic over- or underestimation. The authors are candid about the limits — the models rest on one greenhouse, one season, and mostly sunny conditions, and cross-variety, cross-season generalization remains untested. Still, the demonstration marks a meaningful step toward greenhouses where fleets of small robots continuously read the physiological pulse of their crops, catching stress while it is still invisible to the human eye.

Subject of Research: Autonomous mobile multispectral sensing for non-destructive assessment of greenhouse tomato seedling physiological status

Article Title: An autonomous mobile multispectral sensing system for non-destructive assessment of greenhouse tomato seedling physiological status

Article References: Zhang, H., Liu, S., Xu, P., Ma, Z., Zhang, S., Ma, D., & Yun, Y. (2026). An autonomous mobile multispectral sensing system for non-destructive assessment of greenhouse tomato seedling physiological status. Smart Agricultural Technology, 15, Article 102568. https://doi.org/10.1016/j.atech.2026.102568

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102568

Keywords: multispectral imaging, greenhouse robotics, tomato seedlings, precision agriculture, XGBoost, LiDAR navigation, SLAM, SPAD, leaf nitrogen, non-destructive sensing, NDVI, smart greenhouse

A research team led by Ashmita Upadhyay and Samuel B. Fernandes at the University of Arkansas, together with collaborators at the University of Illinois Urbana-Champaign and Stanford University, hypothesized that the vast, largely untapped reservoir of information in hyperspectral reflectance data could be converted into powerful secondary traits for multi-trait genomic prediction. Rather than relying on well-known vegetation indices such as NDVI, which use only a handful of predefined wavebands, the team generated more than 4.6 million candidate wavelength ratios from a 350 to 2500 nanometer spectral window and screened them statistically for usefulness. Their key insight was that a secondary trait does not need to have any intrinsic biological meaning at all. It simply needs to be highly heritable and strongly genetically correlated with the target trait, a combination captured by a single quantity known as co-heritability.

The case study focused on two physiologically important leaf traits in field-grown sorghum: leaf nitrogen content per unit area and specific leaf area, the lamina area per unit of dry mass. Leaf nitrogen governs photosynthetic capacity, metabolism, litter decomposition, and fertilizer management, while specific leaf area is a cornerstone of the worldwide leaf economics spectrum, linked to growth strategy, defense, and resource investment. Measuring these traits directly is labor-intensive, requiring leaf excision, drying, grinding, and elemental analysis, which limits the scale at which breeders can collect them. The team also included partial least squares regression predictions of both traits from spectral data as additional targets, producing four target traits in total, each with moderate heritabilities between 0.61 and 0.68.

The experimental foundation was a 2017 field experiment with 869 photoperiod-sensitive sorghum lines grown at two University of Illinois research farms in an augmented block design. Hyperspectral reflectance was measured on the youngest fully expanded leaves using a full-range spectroradiometer fitted with an illuminated leaf clip, with careful splice correction and quality control applied to every spectrum. Leaf discs of known area were oven-dried to calculate specific leaf area, and nitrogen content was determined with an elemental analyzer. Genotypic data consisted of roughly 454,000 high-quality SNP markers scored across 836 sorghum lines, imputed and filtered from a larger whole-genome resequencing reference panel.

The computational heart of the study was a massive screening pipeline. For each target trait, the researchers fitted linear mixed models to estimate variance components for every possible ratio of two wavelengths, yielding 4,624,650 co-heritability estimates per trait. The calculation, implemented in a new R package called SyntheticTraits, took approximately four hours per trait on a high-performance computing cluster using 2,112 CPU cores with modest memory requirements per core. From the top one percent of wavelength ratios ranked by co-heritability, hierarchical clustering was used to select three diverse synthetic traits per target trait, giving twelve synthetic traits in all. Their heritabilities ranged from 0.46 to 0.70, their absolute genetic correlations with the target traits ranged from 0.71 to 0.99, and their co-heritabilities ranged from 0.50 to 0.63.

The predictive payoff was evaluated with fivefold cross-validation under three schemes: a single-trait baseline, a CV1 scheme in which both the target trait and the synthetic trait were masked in the validation set, and a CV2 or trait-assisted scheme in which the synthetic trait remained observed for all genotypes. Training was always performed at one farm and validation at the other, a directional transfer design that guards against over-optimistic performance estimates. The results were consistent across all four target traits. Predictive ability rose from 0.29 to 0.34 for leaf nitrogen, from 0.37 to 0.42 for specific leaf area, from 0.41 to 0.45 for PLSR-predicted nitrogen, and from 0.34 to 0.37 for PLSR-predicted specific leaf area, representing gains of up to 17.24 percent over single-trait models. The CV2 scheme, which exploits full spectral coverage of the validation population, was consistently the best performer, while CV1 performed similarly to the single-trait baseline.

Crucially, the researchers included a control that validated their entire selection logic. When they deliberately selected a synthetic trait with a co-heritability of zero, meaning near-zero genetic correlation with the target trait, the multi-trait model offered no improvement whatsoever over the single-trait model, regardless of that synthetic trait’s own heritability. This negative result is as informative as the positive one: it demonstrates that co-heritability is the quantity that matters, not the mere inclusion of any spectral derivative. It also echoes earlier findings that biologically meaningful secondary traits fail to help when their heritability or genetic correlation is low.

An additional practical concern was whether synthetic traits must be identified on the same individuals used to train the prediction models, which could inflate performance through circularity. To test this, the team split the population into five mutually exclusive 20 percent subsets, identified synthetic traits within each subset, and trained genomic prediction models on the remaining 80 percent. The trend held: predictive ability increased by roughly 15.61 percent for leaf nitrogen, 14.24 percent for specific leaf area, 7.44 percent for PLSR-predicted nitrogen, and 6.98 percent for PLSR-predicted specific leaf area. Although absolute accuracies were slightly lower than with the full dataset, the comparison across models was preserved, showing that only a fraction of the population is needed to mine useful spectral ratios. Intriguingly, one particular subset of genotypes consistently produced better models than others, underscoring the well-known influence of training population composition on genomic prediction outcomes.

Although the winning synthetic traits carry no intrinsic biological meaning by construction, many of the wavelengths involved overlap with regions used in established vegetation indices. Near-infrared wavebands around 808 to 811 nanometers, common to indices such as NDVI, EVI, and GNDVI, appeared in synthetic traits for leaf nitrogen and specific leaf area, and red wavebands near 743 nanometers also featured prominently. This suggests the data-driven search is rediscovering genuinely informative parts of the spectrum while also exploiting combinations that no predefined index would capture. The authors note that PLSR-derived target traits may have enjoyed somewhat optimistic results because their predictions were built from the same spectral data, but the consistent gains under CV2 across both direct and PLSR targets support the broader conclusion.

The implications extend well beyond sorghum leaves. The method is general in nature and should transfer to any multidimensional dataset collected alongside target traits, including drone-based remote sensing imagery, controlled-environment phenotyping facilities, microscopy images, metabolomics, and ionomics. Because hyperspectral measurements are rapid and non-destructive, once spectra are collected they can support prediction for many target traits without additional measurement effort. The researchers caution that some improvements were modest and that the approach has so far been demonstrated on point measurements rather than images or time series, and they call for further work integrating field-level high-throughput phenotyping into multi-environment multi-trait models. Still, the demonstration that synthetic traits with no biological meaning, selected purely through co-heritability, can outperform conventional single-trait prediction opens a fundamentally new route to squeezing more value from the deluge of phenomic data flowing through modern breeding programs. The hyperspectral data and code, including the SyntheticTraits R package, have been made publicly available so that breeding teams worldwide can test whether light’s hidden ratios can sharpen their own predictions.

Subject of Research: Using co-heritability-selected synthetic traits from hyperspectral reflectance data to improve multi-trait genomic prediction of leaf nitrogen content and specific leaf area in sorghum

Article Title: Improving multi-trait genomic prediction using synthetic traits from hyperspectral data based on co-heritability

Article References: Upadhyay, A., Azam, R., Yuan, M., Ivanovic, S., Ferguson, J. N., Paul, R. E., Koyejo, S., El-Kebir, M., Lipka, A. E., Leakey, A. D. B., & Fernandes, S. B. (2026). Improving multi-trait genomic prediction using synthetic traits from hyperspectral data based on co-heritability. Theoretical and Applied Genetics, 139(10), Article 270. https://doi.org/10.1007/s00122-026-05334-2

Image Credits: AI Generated

DOI: 10.1007/s00122-026-05334-2

Keywords: genomic prediction, hyperspectral data, synthetic traits, co-heritability, multi-trait models, sorghum, leaf nitrogen, specific leaf area, high-throughput phenotyping, GBLUP, plant breeding, wavelength ratios