Hyperspectral imaging predicts plant-based meat texture without destroying the sample

Plant-based meat analogs have moved from novelty to mainstream, but one stubborn bottleneck remains: texture. The bite of a burger patty, its chew, its spring back against the teeth—these are the qualities that decide whether a consumer comes back for a second taste. The problem is that measuring texture has always meant destroying the product. A metal probe compresses the sample, the sample never recovers, and every measurement consumes a specimen that took time, ingredients, and energy to make. Now a team of South Korean researchers has demonstrated a way to read the texture of plant-based meat analogs without touching them at all, using nothing more than light and a carefully trained statistical model.

The study, published in Food Science and Biotechnology, was led by Hyun-Su Choi and Jaeeun Lee of Seoul National University, working with colleagues across the Department of Biosystems Engineering and the Research Institute of Agriculture and Life Sciences. Their approach combines hyperspectral imaging with partial least squares regression, a chemometric workhorse, to predict six textural attributes of model plant-based meat analogs formulated with sesame seed meal protein isolate. The work is notable not only for the imaging technology but also for the ingredient itself: sesame seed meal, a by-product left over after sesame oil extraction, contains roughly 35 to 40 percent crude protein and has been largely overlooked as a meat analog building block.

To make the protein ingredient, the researchers took sesame seed meal provided by a cold-press oil producer and put it through a multi-step extraction. The meal was dispersed in water, homogenized at 12,000 rpm, heated at 80 degrees Celsius for two hours, and then stirred overnight. After centrifugation removed the oil layer, the pH was raised to 10.0 to solubilize the proteins, then dropped to 4.0 to precipitate them. The recovered protein was neutralized, frozen at minus 80 degrees Celsius, and freeze-dried for five days. The resulting sesame seed meal protein isolate powder contained 77.80 percent crude protein, with a yield of 19.52 percent from the original meal.

With the isolate in hand, the team prepared 81 model meat analogs by mixing the protein with water at concentrations between 25 and 35 percent by weight. The mixtures were hydrated, poured into Petri dishes, sealed, and heat-treated at 90 degrees Celsius for 20 minutes. The samples spanned a wide textural range: group-mean hardness varied from about 5,462 to 38,389 grams, and chewiness from about 3,354 to 29,246 grams. Samples with lower moisture content tended to be harder and chewier, consistent with earlier reports that higher moisture softens plant protein matrices. Cohesiveness and resilience also differed among groups, while springiness stayed remarkably constant, varying only between 0.89 and 0.95.

The imaging system was a push-broom line-scan hyperspectral camera covering 900 to 1,700 nanometers across 436 bands, illuminated by a 600-watt halogen source. Each sample was scanned as the camera moved along a motorized stage, with the optical axis perpendicular to the surface at a fixed distance of 300 millimeters. Reflectance was calibrated using a standard white panel with 95 percent reflectance and a dark reference image. Individual sample regions were segmented by thresholding at 1,070 nanometers, a wavelength where sample reflectance is near its maximum and contrast against the background is largest. The team confirmed that segmenting at the central band of 1,272 nanometers produced identical regions and mean spectra correlated at r greater than 0.99999, showing that the choice of segmentation band did not affect the extracted data.

The reflectance spectra of all 80 samples showed similar overall patterns, with absorption troughs at the O–H second overtone near 970 nanometers, the C–H second overtone near 1,200 nanometers, and the O–H first overtone near 1,450 nanometers. Between-sample variation was relatively large in the 1,000–1,100 and 1,200–1,400 nanometer regions, while reflectance dropped to around 0.1 in the 1,400–1,500 nanometer region, where the strong water absorption near 1,450 nanometers suppressed the reflected signal. No samples were flagged as spectral outliers by Hotelling’s T-squared and Q residual tests, so all 80 were used for model development.

Partial least squares regression models built from the mean spectra showed markedly different performance across the six texture attributes. Hardness came out on top, with a cross-validation coefficient of determination of 0.882 and a residual predictive deviation of 2.92. Resilience followed at 0.865 and 2.74, and chewiness at 0.829 and 2.44. By the conventional interpretation, residual predictive deviation values between 2 and 3 indicate discrimination—adequate for screening but not for fully quantitative prediction. Cohesiveness showed intermediate performance at 0.669 and 1.75, adhesiveness was weaker at 0.607 and 1.61, and no valid model could be established for springiness, whose coefficient of variation was only 4.5 percent, leaving too little variation for the model to explain.

The team also examined why the models performed as they did. Hardness, chewiness, and resilience were mutually correlated at r greater than or equal to 0.96, reflecting a shared compression-resistance axis tied to moisture state and structural density—properties well represented in near-infrared reflectance. Adhesiveness, in contrast, correlated only weakly at 0.49 and reflects surface tackiness governed by local surface state rather than bulk composition. The achievable performance was also bounded by the reproducibility of the reference texture measurements themselves: within-group scatter accounted for only 7 to 10 percent of total variance for hardness, chewiness, and resilience, setting upper bounds of R-squared 0.90 to 0.93, which the models achieved 89 to 97 percent of. For cohesiveness and adhesiveness, within-group scatter was 16 and 20 percent, lowering the bounds to 0.84 and 0.80.

Variable importance in projection analysis identified the wavelength regions driving the predictions. The three attributes with adequate performance—hardness, resilience, and chewiness—shared 109 selected bands concentrated in three regions: 900–1,036, 1,134–1,145, and 1,171–1,213 nanometers. Across all five modeled attributes, 85 bands were common, occurring in the regions of 900–933, 948–1,035, 1,134–1,141, and 1,171–1,191 nanometers. These shorter-wavelength regions include the O–H second overtone of water near 970 nanometers and overlapping C–H overtone features. Notably, the region above 1,400 nanometers was not selected for any attribute, even though the first water overtone near 1,450 nanometers shows the strongest absorption—suggesting that information there was largely redundant with shorter wavelengths. Rebuilding models with only the selected wavelengths reduced the variable count to 34 to 41 percent of the total without improving accuracy; for hardness, the band count dropped by 66 percent, from 436 to 150, while the cross-validation R-squared was maintained at 0.862 compared with 0.882.

The findings point toward a practical future for plant-based meat quality control. A reduced-wavelength sensing system built around the 900–1,213 nanometer range could screen formulations rapidly, cutting the cost and time of texture evaluation during product development. The researchers caution that the study was based on 80 samples from a single material system, and that further validation using independent batches and a wider range of materials and processing conditions is needed to assess generalizability. Still, the demonstration that a camera and a regression model can stand in for a destructive texture probe marks a meaningful step toward nondestructive, high-throughput quality assessment for the next generation of plant-based foods.

Subject of Research: Food Science and Biotechnology

Article Title: Nondestructive prediction of textural properties of model plant-based meat analogs prepared with sesame seed meal protein isolate using hyperspectral imaging

Article References: Choi, H.-S., Lee, J., Jung, I.-G., Jang, H., Sung, D.-H., Yoon, P. R., Park, S., Kim, G., & Kim, Y.-R. (2026). Nondestructive prediction of textural properties of model plant-based meat analogs prepared with sesame seed meal protein isolate using hyperspectral imaging. Food Science and Biotechnology. https://doi.org/10.1007/s10068-026-02335-1

Image Credits: AI Generated

DOI: 10.1007/s10068-026-02335-1

Keywords: hyperspectral imaging, plant-based meat analogs, sesame seed meal protein isolate, texture profile analysis, partial least squares regression, nondestructive analysis, Nondestructive, prediction, textural, properties, model, plant-based