Revealing hidden insect damage in sunflower seeds using near-infrared hyperspectral imaging and machine learning

Authors: KAKKAR, SATYAM - North Dakota State University; MENSAH, BRIGHT - North Dakota State University; Prasifka, Jarrad; Hulke, Brent; MONONO, EWUMBUA - North Dakota State University; SUN, XIN - North Dakota State University

Submitted to: Journal of Agriculture and Food Research
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 4/18/2026
Publication Date: 4/21/2026
Citation: Kakkar, S., Mensah, B., Prasifka, J.R., Hulke, B.S., Monono, E., Sun, X. 2026. Revealing hidden insect damage in sunflower seeds using near-infrared hyperspectral imaging and machine learning. Journal of Agriculture and Food Research. 28. https://doi.org/10.1016/j.jafr.2026.102945.
DOI: https://doi.org/10.1016/j.jafr.2026.102945

Interpretive Summary: Estimating the amount of insect damage in seed samples is an important task. For industry, it ensures that the food going to consumers meets standards of taste and appearance. For researchers, estimates of insect damage help show if management of a pest has been successful, especially for testing new pest management methods. The problem is that estimating insect damage in seeds can be slow and costly. In this study, wavelengths of light beyond what human eyes can see (near-infrared or NIR) were used to test if insect damage estimates from NIR are similar to human inspection of sunflower seed samples. Several different methods for analyzing the information from NIR wavelengths were also tested. The varied methods of analysis were accurate from 80% to 90% of seeds. The best method appears suitable for sunflower seeds, but additional work is needed to add the technology into large (industrial) or small (research) systems that handle and analyze seed in a single unit.

Technical Abstract: Reliable detection of insect-infected sunflower seeds is critical for maintaining seed quality and minimizing economic losses in commercial processing chains. Hyperspectral Imaging (HSI) in the near-infrared region offers a non-destructive way to capture subtle biochemical and structural changes caused by insect feeding, but effective exploitation of its high-dimensional data remains challenging. This study developed a complete seed-level HSI pipeline in the range of 900-1700 nm to identify damaged and undamaged seeds based on their spectral characteristics. In this study, two spectral feature representations were evaluated: Principal Component Analysis feature set and a hybrid feature space combining PCA with eight continuum-removal (CR) metrics describing the depth, area, centroid, and width of moisture and lipid related absorption bands. Using these features, 6 machine learning classifiers namely, Multi-Layer Perceptron (MLP), Support Vector Machine (SVM), Random Forest (RF), LightGBM, XGBoost, and PLS-DA, were trained with 5-fold cross-validation and evaluated on an hold-out test set. With PCA-only features, XGBoost and MLP achieved accuracies of 83.53% and 82.93%, respectively. Incorporating CR features improved performance, with the MLP reaching 88.11% accuracy. A soft-ensemble of the optimized MLP and SVM models yielded the best performance, achieving 89.98% accuracy. These results demonstrate that combining dimensionality reduction with CR-based descriptors substantially enhances class separability and provides a practical pathway toward scalable, HSI-based industrial seed inspection systems.