Can Infrared Scans Expose Added Syrup in Bangladeshi Honey?
A study of four honey varieties from Bangladesh tested whether infrared chemical patterns and machine learning could spot laboratory-added corn syrup or caramel color. The reported results were strong, but the method remains a regional proof of concept.
The full story
What Happened?
Researchers studied whether an infrared “fingerprint” of honey could help a computer recognize botanical variety, distinguish pure from adulterated samples, and identify the added adulterant. [2] [1]
They assembled a reported dataset of 1,040 samples representing Blackseed, Litchi, Mustard, and Rubber honeys from Bangladesh. Samples were prepared with corn syrup or caramel color at concentrations from 2.5% to 30%. Six machine-learning models were compared using information collected by Fourier-transform infrared spectroscopy, or FTIR. [2] [1]
Why Should We Care?
Honey fraud can be difficult to investigate when added ingredients resemble honey’s natural sugar mixture. This study explores a rapid, non-destructive screening approach that reads chemical patterns without consuming the sample. [2]
The results suggest that FTIR combined with machine learning could become a useful authentication tool. However, the work is best understood as a regional proof of concept—not yet a universally validated test for all honeys or adulterants. [2]
What Did the Researchers Find?
- The reported dataset contained 1,040 samples representing four honey types: Blackseed, Litchi, Mustard, and Rubber. [2] [1]
- Researchers prepared samples with corn syrup or caramel color at 12 concentrations between 2.5% and 30%. [2] [1]
- The system addressed three tasks: botanical-variety classification, pure-versus-adulterated classification, and identification of the added adulterant. [2]
- Feature selection produced a separate set of 15 diagnostic spectral measurements for each classification task; many important measurements occurred in the carbohydrate fingerprint region. [2] [1]
- Random Forest, CatBoost, and K-nearest neighbors reportedly reached 100% accuracy, precision, and F1 scores under the study’s evaluation conditions; the linear support-vector machine was the weakest model, with reported accuracy as low as 92.31%. [2] [1]
- A spectral measurement at 1740.7 cm⁻¹ emerged as the leading indicator in the study’s pure-versus-adulterated classification task. [2] [1]
How Do We Know?
The researchers collected FTIR spectra from pure and deliberately adulterated honey samples. FTIR records how a sample absorbs infrared light across different wavelengths, producing a pattern related to its chemical composition. [2] [1]
They compared six machine-learning approaches: Random Forest, linear support-vector machine, LightGBM, XGBoost, CatBoost, and K-nearest neighbors. Recursive feature elimination with cross-validation was used to reduce the spectra to selected diagnostic measurements, while principal component analysis was used to examine clustering. [2] [1]
The accessible abstract and publisher-indexed text report very high classification performance. The complete validation workflow, data partitioning, and model-independence details could not be fully checked without direct full-text access. [2] [1]
What Doesn’t This Study Prove?
Regional evidence The pure honeys represented four botanical varieties sourced in Bangladesh. The authors describe the work as a regional proof of concept, so its spectral markers and model boundaries may not transfer unchanged to honeys from other climates, floral sources, or regions. [2]
Only two adulterants tested The experiment examined corn syrup and caramel color. It does not prove that the models can recognize other syrups, feeding-related adulteration, botanical mislabeling outside the four studied varieties, or unfamiliar mixtures. [2] [1]
Laboratory-prepared samples The adulterated samples were systematically prepared at known concentrations. High performance under these controlled conditions does not establish equal accuracy on unknown retail samples affected by storage, processing, harvest year, moisture, or multiple adulterants. [2]
Detection below 2.5% remains unproven The lowest tested adulteration concentration was 2.5%. The study therefore does not demonstrate reliable detection at smaller concentrations. [2] [1]
Perfect scores are not universal guarantees The reported 100% results describe particular models evaluated on this study’s dataset. They should not be presented as guaranteed performance in laboratories, inspections, or commercial supply chains without independent and external validation. [2] [1]
Where Did This Research Come From?
Research Record
- Original Study: FTIR-driven multi-task learning for non-destructive honey adulteration detection in Bangladesh
- Authors: Alkatuzzakia Akhi; Md Abdullah Al Noman; Iqbal Hossain; Anannya Barua Nijhum; Ms Asmaul Hosna Ema; Mohammad Gulzarul Aziz; Afzal Rahman
- Journal: Food Control
- Publisher: Elsevier BV
- Publication: 2027-01 journal issue; exact online publication date not verified
- Honey Chronicles Timeline Date: Exact timeline date needs verification
- Journal Issue Date: 2027-01
- Volume: 191
- Pages / Article: 112452
- Research Location / Affiliations: Honey samples sourced in Bangladesh; author affiliations not verified from the accessible scholarly records
- DOI: 10.1016/j.foodcont.2026.112452