Multimodal Approach to Quinoa Genotype Classification: Combining Image Analysis and Near-Infrared Spectroscopy
NEW ZEALAND JOURNAL OF CROP AND HORTICULTURAL SCIENCE, cilt.54, ss.1-19, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 54
- Basım Tarihi: 2026
- Doi Numarası: 10.1002/nzc2.70204
- Dergi Adı: NEW ZEALAND JOURNAL OF CROP AND HORTICULTURAL SCIENCE
- Derginin Tarandığı İndeksler: Academic Search Ultimate (EBSCO), Natural Science Collection (ProQuest), Biomedical Reference Collection: Corporate Edition (EBSCO), Engineering Source (EBSCO), Scopus, Science Citation Index Expanded (SCI-EXPANDED), Periodicals Index Online, BIOSIS, Geobase
- Sayfa Sayıları: ss.1-19
- Çanakkale Onsekiz Mart Üniversitesi Adresli: Evet
Özet
Distinct genotypic differences in the phenotypic and chemical composition of quinoa (Chenopodium quinoa Willd.) seeds complicate standardization and traceability processes, and methods based solely on morphological analysis are often insufficient. This study integrated (i) physical-morphometric characteristics, (ii) RGB color attributes, and (iii) the chemically predicted composition via the near-infrared (NIR) spectrum (1200–2400 nm) to classify three quinoa genotypes (CHN1, CHN2, CHN3). The study also compared the performance of feature-based machine learning algorithms and transfer learning-based CNN architectures using raw images. Morphometric characteristics and average RGB color values were derived from seed images obtained at a resolution of 1200 dpi under controlled conditions. Using the NIR spectrum, crude ash, total carbohydrate, total starch, crude protein, crude fat, crude fiber, total soluble sugar, and moisture contents of the seeds were estimated. It was determined that the genotypes showed significant differences both phenotypically and compositionally (p < 0.05). Both feature-based machine learning approaches and transfer learning-based CNN architectures successfully distinguished quinoa genotypes, demonstrating the effectiveness of complementary analytical strategies. Machine learning algorithms (CHAID, NB, SMO, RF, MLP, REPTree) using the combined feature set distinguished the genotypes with high accuracy by class-based and overall model metrics (90%–95.5%). The CHAID algorithm demonstrated the best overall performance with an accuracy rate of 93.8%.In addition to machine learning approaches, four different transfer learning-based CNN architectures (Xception, DenseNet121, ResNet50, InceptionResNetV2) were used. Xception, DenseNet121, and InceptionResNetV2 achieved high results with reaching 95.6% overall classification accuracy, outperforming ResNet50. As a result, the multimodal integration of visual phenotype and NIR-based compositional data enabled accurate and scalable classification of quinoa genotypes, supporting seed purity control and genotype-based traceability.