Hyperspectral imaging in process analytics: what the 2025-2026 reviews say

Hyperspectral imaging (HSI) has moved from a remote-sensing curiosity into a routine entry on conference programmes for pharmaceutical, food, and waste-processing audiences. The last twelve months have produced an unusually dense crop of review papers across those domains. Read together, they make a more useful argument about where the technique sits than any one of them does alone.

This is a reading list, not a vendor scorecard. The six papers below were selected from recent issues of Advanced Science, Microchemical Journal, the Journal of Agricultural and Food Chemistry, Analytical Chemistry, Technologies, and Waste Management. They cover separate application stacks, but they keep landing on the same three structural constraints: data analysis is still the bottleneck, deep learning helps when there is enough labelled data, and the field still lacks shared protocols. Readers new to the multivariate side can start with our chemometrics 101 primer; readers tracking the deep-learning angle on a sibling modality can consult our deep-learning chemometrics for process Raman note.

Lyophilization: the most ambitious pharma review

The most substantial pharmaceutical-side paper is Yu, Srisuma, Devos, Wang, Myerson and Braatz, “Hyperspectral Imaging Techniques for Lyophilization: Advances in Data-Driven Modeling Strategies and Applications” (Advanced Science 12(33):e08506, 23 July 2025, DOI 10.1002/advs.202508506). The MIT Braatz group surveys HSI applications in freeze-drying across residual moisture mapping, solid-state transformations, phase separation in protein formulations, and particulate detection (the authors cite work resolving 2 mm particulates), with prediction errors as low as 0.15 in the reviewed regression studies.

What separates the paper from older HSI surveys is its data-pipeline framing. The authors systematise preprocessing, spectral unmixing, classification and regression, and data fusion, and they include emerging tensor-based and deep-learning methods alongside the conventional PLS chemometrics. Their explicit framing is quality-by-design rather than quality-by-testing, with continuous manufacturing and complex biopharmaceuticals (the paper names mRNA vaccines) as the motivating applications. For teams considering HSI for a critical step like lyophilization, this is now the natural literature anchor; for the wider PAT context, the review is also one of the cleaner restatements of why imaging beats point spectroscopy when spatial heterogeneity is itself a quality attribute.

Food data analysis: critical-review honesty

The Microchemical Journal review by the Universite de Lille group, “Strategies for analysing hyperspectral imaging data for food quality and safety issues - A critical review of the last 5 years” (May 2025, S0026265X25013487), takes the opposite tack. It is unconcerned with hardware and instead audits how five years of food-HSI papers handled their data. The split the authors highlight is between treating the sample as a whole versus treating it as a set of pixel-spectra; the choice cascades through preprocessing, model selection, and validation strategy.

PCA, PLS regression, PLS-DA, SIMCA, and SVM dominate the literature regardless of whether the sample is plant- or animal-based. The authors are unusually direct about limits: high instrument cost, computational load, and the absence of standardised protocols. Anyone planning a food-HSI deployment will find the data-strategy decision tree more useful than the application gallery.

Fruits, vegetables, and deep learning

Yang, Guo, Barbin, Dai, Watson, Povey and Zou’s “Hyperspectral Imaging and Deep Learning for Quality and Safety Inspection of Fruits and Vegetables: A Review” (Journal of Agricultural and Food Chemistry 73(17):10019-10035, 30 April 2025, DOI 10.1021/acs.jafc.4c11492) is the deep-learning counterpart. It catalogues HSI-plus-CNN work across external defects, internal quality (sugar, firmness, ripeness), and pathogen detection. The pattern is consistent with what is being reported on the spectroscopy side: deep models outperform classical chemometrics where labelled spectra are plentiful, and lose or tie where they are not. The transferable lesson for industrial users is that the architecture choice is downstream of the labelling budget.

Plant pigments in food: a 2026 entry

The newest paper on the list is Zhang and colleagues’ “Applications and Advances in Hyperspectral Imaging for Detecting Three Major Natural Plant Pigments in Food: A Review” (Analytical Chemistry 98(5):3387-3412, 29 January 2026, DOI 10.1021/acs.analchem.5c03972). The Jiangsu University group concentrates on carotenoids, anthocyanins, and chlorophylls as both quality markers and authenticity indicators. The pigment focus matters because these compounds have well-characterised visible/NIR signatures, which makes them ideal stress tests for whether HSI plus modern chemometrics is reaching the precision needed for compositional assay rather than classification.

Plastics sorting: an industrial benchmark

For an applied counterweight, the Waste Management paper “Plastics detection and sorting using hyperspectral sensing and machine learning algorithms” (S0956053X2500265X, May 2025) evaluates HSI across 8 polymer typologies (PS, PC, PE-types, PP, PVC, PET, PLA, and MaterBi) in the 900-1650 nm range. The technical headline is that machine-learning classifiers on HSI data hit accuracies that justify line-speed deployment for most of those polymers; the practical headline is that black plastics remain the unresolved case because carbon black absorbs the NIR band, which is why mid-wave-infrared HSI is moving into the same conversation. For process-analytics readers outside recycling, the value of the paper is that it documents how an HSI deployment actually performs across a heterogeneous feedstock at industrial throughput, which is rarer than it should be.

The cross-disciplinary view

Cheng, Mukundan, Karmakar, Valappil, Jouhar and Wang’s “Modern Trends and Recent Applications of Hyperspectral Imaging: A Review” (Technologies 13(5):170, 23 April 2025) is the one to send to a manager who wants the wide angle. The paper proposes a cross-disciplinary classification across medical, agricultural, environmental, and industrial uses, and its strongest contribution is structural rather than technical: it shows how thinly the same algorithmic toolbox is being spread across very different application contexts, and how often application-specific protocols are missing.

Three threads run through the set. First, the bottleneck is no longer the camera; sensor cost is still high but is no longer the binding constraint on whether HSI works for a given task. Second, deep learning is being adopted in HSI on a similar schedule to spectroscopy, with the same caveat that data-rich settings favour it and data-poor ones do not. Third, none of the reviews can point to mature standards for preprocessing, validation, and reporting, which is precisely the gap that holds HSI back from the kind of regulatory-grade deployment that point-spectroscopy methods have started to achieve. For pharmaceutical readers, the lyophilization review and the Yu et al. data-pipeline framing are the practical entry points; for food and recycling readers, the Microchemical Journal and Waste Management papers are the more concrete reference. None of the six retires conventional chemometrics. Together they make a credible case that HSI’s next round of method development should at least include a deep-learning baseline alongside the PLS one, and that the model-validation work needed before any of this reaches a GMP environment is still ahead of the literature, not behind it.