Location: Quality & Safety Assessment ResearchTitle: Near-infrared Hyperspectral Reflectance Imaging for Early Detection of Sour Skin Disease in Vidalia Sweet Onions
|WANG, WEILIN - University Of Georgia|
|LI, CHANGYING - University Of Georgia|
|GITAITIS, RON - University Of Georgia|
|TOLLNER, E.W. - University Of Georgia|
|RAINS, GLEN - University Of Georgia|
Submitted to: ASABE Annual International Meeting
Publication Type: Proceedings
Publication Acceptance Date: 11/20/2010
Publication Date: 11/20/2010
Citation: Wang, W., Li, C., Gitaitis, R., Tollner, E., Rains, G., Yoon, S.C. 2010. Near-infrared Hyperspectral Reflectance Imaging for Early Detection of Sour Skin Disease in Vidalia Sweet Onions. ASABE Annual International Meeting. Paper No. 1009106.
Technical Abstract: Sour skin is a major onion disease caused by the bacterium Burkholderia cepacia (B. cepacia). It not only causes substantial economic loss from diseased onions but also could lead to pulmonary infection in humans. It is critical to prevent onions infected by sour skin from entering storage rooms or being shipped to fresh vegetable markets. This paper reports the development of a hyperspectral imaging method for early detection of onions infected by sour skin. In this study, near-infrared hyperspectral reflectance images of 40 Vidalia sweet onions were taken in 2 nm increments from 950 nm to 1650 nm, before and after they were inoculated with B. cepacia. Inoculated onion samples were scanned every day after inoculation for 7 days, while the hyperspectral images scanned before inoculation were used as controls. Spectral signatures of onion hyperspectral images were extracted from selected regions of interest. Based on the principal component analysis conducted on spectral signatures of control and inoculated samples, two optimal spectral bands (1070nm and 1400nm) were selected to construct ratio images, which better revealed the difference between the control and inoculated samples. Mean ratio values at three different areas on the onion surface (flesh body area, root or neck area, and the whole onion area) were calculated from ratio images and used as inputs for classification models. The three spatial features of mean ratio values obtained from band-ratio images were proved to be good indicators of sour skin-infected onions. When comparing two classifiers, The back-propagation neural network (BPNN) models performed better (95% accuracy) than support vector machine (SVM) classifiers (85%-90%) in discriminating control samples and inoculated samples on day 6 after inoculation. Then, the optimal BPNN classifier using three spatial features of band-ratio images was applied to classify hyperspectral images of tested onion samples over the period of 1-7 days after inoculation, respectively. The results of tests showed that the near-infrared hyperspectral reflectance imaging technique could detect sour skin-infected onions effectively from day 4 to day 7 after inoculation by achieving overall classification accuracies of 80%, 85%, 95%, and 100%, respectively.