AI-Based Classification of NIR Spectra for Model Hard Candies: Impact of Storage Time
sensAIfood Final International Conference – Application of AI Methods for Processing Spectral Data in Feed and Food Chain, Tirane, Arnavutluk, 8 - 10 Ekim 2025, ss.54, (Özet Bildiri)
- Yayın Türü: Bildiri / Özet Bildiri
- Basıldığı Şehir: Tirane
- Basıldığı Ülke: Arnavutluk
- Sayfa Sayıları: ss.54
- Orta Doğu Teknik Üniversitesi Adresli: Evet
Özet
Hard candies are sugar-based confections whose structural and physicochemical stability can be affected by storage conditions over time. Especially in sugar-free formulations, storage may alter molecular interactions, leading to measurable changes in their spectral properties. In this study, model hard candies were prepared with 92.5% isomalt, 2.0% citric acid, and sorbitol as a sugar alcohol. A total of 240 samples were measured at three different storage times (2, 3, and 4 months). Spectral data were acquired using near-infrared (NIR) hyperspectral imaging (HSI) between 953–2500 nm. After preprocessing using Savitzky–Golay smoothing (11-point window) and specific wavelength selection (1112- 1404 nm), the dataset was divided into training (60%) and test (40%) subsets. Three AI-based classifiers that are k-nearest neighbors (kNN), random forest (RF), and neural network (NN) were applied to assess the impact of storage duration on sample classification. The test set classification accuracies were 0.989 for kNN, 0.979 for RF, and 0.929 for NN. These results demonstrate that NIR–HSI spectroscopy combined with appropriate AI models can effectively detect storage-time-dependent spectral variations in hard candy formulations.