The Latest in AI-Enhanced Food Scanning Technology

A recent review article published in the journal Food Physics discussed how combining near-infrared (NIR) spectroscopy with machine learning (ML) is reshaping how the food industry checks quality, authenticity, and safety.¹ However, the researchers also highlighted that unresolved technical and regulatory gaps are preventing widescale adoption on the factory floor.

What did the researchers do in their study?

The analysis, led by Nayeem Mia with researchers from Gyeongsang National University, Bangladesh Agricultural University, and the University of Tasmania, synthesizes existing research on NIR–machine learning (NIR–ML) systems across five major food categories: meat and fish, dairy, cereals and grains, fruits and vegetables, and processed or fermented foods.^1,2 The main takeaway of the review article is that while these systems have proven their worth in controlled research settings, industrial adoption is being slowed by issues of instrument transferability, model drift, and a lack of standardized validation protocols.¹

NIR spectroscopy works by measuring how food samples absorb NIR light, producing spectral data linked to chemical composition.^1,3,4 Unlike destructive laboratory tests, it can assess a product in seconds without damaging it.^1,4 The technology's appeal for food manufacturers lies in its comparatively low cost, simple instrumentation, and suitability for continuous, high-volume monitoring, which are advantages the review notes over competing nondestructive methods such as hyperspectral imaging, nuclear magnetic resonance, X-ray imaging, and ultrasonic testing.¹

What did the review article discussed?

In their review, the authors examined a few chemometric techniques such as partial least squares regression (PLSR) and linear discriminant analysis (LDA). Comparing PLSR and LDA against the newer ML and deep-learning (DL) approaches, the authors concluded that older, simpler models remain valuable for their interpretability and robustness in quantitative quality control, whereas DL methods offer stronger performance on complex, nonlinear spectral relationships but at higher computational cost and with less transparency.¹

Once the researchers came to this conclusion, the team examined all five food sectors. The review identifies a consistent pattern: a clear progression from small, laboratory-scale feasibility studies toward emerging industrial-scale pilots, but with a substantial gap between the two.¹ Laboratory studies tend to report high predictive accuracy under tightly controlled conditions, the authors note, but often without the external validation, batch-effect testing, or long-term drift monitoring needed to confirm that a model will perform reliably on a factory line over time or across different instruments.¹

What are the remaining obstacles in the deployment of food scanning technology?

Food scanning technology currently is facing several barriers that is preventing its wider deployment in the industry. These obstacles include data scarcity, inconsistent preprocessing and acquisition protocols, sensor stability over time, and the "black box" nature of many advanced ML models.¹ The authors argue that preprocessing choices, algorithm selection, and validation design cannot be treated as generic steps; they must be tailored to the specific food matrix and application to produce results that hold up outside the laboratory.¹

What is next in the development of NIR sensors?

The development of NIR sensors in the future is set to focus on portable and lower-cost sensors. It is also expected that inline and online process analytical technology (PAT) will also be a driver of broader adoption of NIR sensors.¹ The authors frame standardization of validation workflows and development of more interpretable models as prerequisites for regulatory acceptance, rather than optional refinements.¹

So what does this mean for food processors? The main takeaway is that NIR–ML tools are improving rapidly and as a result are becoming effective tools for many applications, but they should still be used cautiously as limitations remain.

“Addressing these limitations will require coordinated efforts in data management, standardization of acquisition and validation workflows, and the development of interpretable and transparent modeling approaches that align with regulatory expectations,” the authors wrote in their study.¹

References

1. Mia, N.; Hashem, M. A.; Halim, M. A.; et al. Application of NIR Spectroscopy with Machine Learning in the Food Industry: A Comprehensive Review. Food Phys. 2026, 3, 100089. DOI: 10.1016/j.foodp.2026.100089
2. Wetzel, W. Advancing NIR and Imaging Spectroscopy in Food and Bioanalysis. Spectroscopy Online, 2026. https://www.spectroscopyonline.com/view/advancing-nir-and-imaging-spectroscopy-in-food-bioanalysis (accessed July 7, 2026).
3. Dai, L.; Luo, D.; Zhang, J.; et al. Near-Infrared Spectroscopy in Food Analysis: Applications, Chemometric Strategies, and Technological Advances. Foods 2026, 15 (10), 1814. DOI: 10.3390/foods15101814
4. Fodor, M.; Matkovits, A.; Benes, E. L.; Jokai, Z. The Role of Near-Infrared Spectroscopy in Food Quality Assurance: A Review of the Past Two Decades. Foods 2024, 13 (21), 3501. DOI: 10.3390/foods13213501