{"title":"Nondestructive Identification of Wheat Species using Deep Convolutional Networks with Oversampling Strategies on Near-Infrared Hyperspectral Imagery","authors":"Nitin Tyagi, Sarvagya Porwal, Pradeep Singh, Balasubramanian Raman, Neerja Garg","doi":"10.1007/s10921-024-01143-z","DOIUrl":null,"url":null,"abstract":"<div><p>Differentiating between wheat species poses a significant challenge to the Indian grain industry. Visual inspection of wheat species has drawbacks, including inconsistency, low throughput, and labor intensiveness. In this study, near-infrared hyperspectral imaging (NIR-HSI) was utilized in conjunction with a deep learning approach to achieve precise predictions of wheat at the species level. A dataset comprising 40 different varieties from four Indian wheat species, namely <i>Triticum aestivum</i> (<i>T. aestivum</i>), <i>Triticum durum</i> (<i>T. durum</i>), <i>Triticum dicocccum</i> (<i>T. dicoccum</i>), and <i>Triticale</i>, was prepared using a NIR-HSI system that encompassed the wavelength ranging from 900–1700 nm. The imbalanced dataset is a common problem in the classification task, making it harder for the classifier to classify minority class data correctly. To address this issue, oversampling techniques such as Synthetic Minority Oversampling Technique (SMOTE) and Adaptive Synthetic sampling (ADASYN) were employed. For the classification task, a 1D Convolutional Neural Network (1D-CNN), a 1D-ResNet, and four traditional machine learning models: Naive Bayes (NB), K-Nearest Neighbor (KNN), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) are utilized and compared. The performance of these models was assessed using both imbalanced and balanced datasets. The 1D-CNN model outperformed traditional machine learning models, achieving impressive test accuracies of 98.25% and 98.43% with the SMOTE and ADASYN approach, respectively. These findings underscore the efficacy of NIR-HSI in conjunction with an end-to-end 1D-CNN and oversampling techniques as a reliable and efficient method for the rapid, accurate, and nondestructive identification of various wheat species. The code is available at https://github.com/nitintyagi007-iitr/Wheat_species_classification</p></div>","PeriodicalId":655,"journal":{"name":"Journal of Nondestructive Evaluation","volume":"44 1","pages":""},"PeriodicalIF":2.6000,"publicationDate":"2024-12-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Journal of Nondestructive Evaluation","FirstCategoryId":"88","ListUrlMain":"https://link.springer.com/article/10.1007/s10921-024-01143-z","RegionNum":3,"RegionCategory":"材料科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"MATERIALS SCIENCE, CHARACTERIZATION & TESTING","Score":null,"Total":0}
引用次数: 0
Abstract
Differentiating between wheat species poses a significant challenge to the Indian grain industry. Visual inspection of wheat species has drawbacks, including inconsistency, low throughput, and labor intensiveness. In this study, near-infrared hyperspectral imaging (NIR-HSI) was utilized in conjunction with a deep learning approach to achieve precise predictions of wheat at the species level. A dataset comprising 40 different varieties from four Indian wheat species, namely Triticum aestivum (T. aestivum), Triticum durum (T. durum), Triticum dicocccum (T. dicoccum), and Triticale, was prepared using a NIR-HSI system that encompassed the wavelength ranging from 900–1700 nm. The imbalanced dataset is a common problem in the classification task, making it harder for the classifier to classify minority class data correctly. To address this issue, oversampling techniques such as Synthetic Minority Oversampling Technique (SMOTE) and Adaptive Synthetic sampling (ADASYN) were employed. For the classification task, a 1D Convolutional Neural Network (1D-CNN), a 1D-ResNet, and four traditional machine learning models: Naive Bayes (NB), K-Nearest Neighbor (KNN), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) are utilized and compared. The performance of these models was assessed using both imbalanced and balanced datasets. The 1D-CNN model outperformed traditional machine learning models, achieving impressive test accuracies of 98.25% and 98.43% with the SMOTE and ADASYN approach, respectively. These findings underscore the efficacy of NIR-HSI in conjunction with an end-to-end 1D-CNN and oversampling techniques as a reliable and efficient method for the rapid, accurate, and nondestructive identification of various wheat species. The code is available at https://github.com/nitintyagi007-iitr/Wheat_species_classification
期刊介绍:
Journal of Nondestructive Evaluation provides a forum for the broad range of scientific and engineering activities involved in developing a quantitative nondestructive evaluation (NDE) capability. This interdisciplinary journal publishes papers on the development of new equipment, analyses, and approaches to nondestructive measurements.