Articles
Investigating spectral imaging for predicting tomato sepal sensitivity of recently harvested tomatoes to fungal infections
Article number
1396_14
Pages
99 – 106
Language
English
Abstract
We address the problem of identifying tomato sepals in recently harvested tomatoes which are susceptible to fungal infections.
Near-infrared (NIR) spectral imaging (1000-1700 nm) is explored for identification of the susceptible tomato sepals before any visual signs of infections appear.
Spectral images are acquired for the tomato cultivar Cappricia, a day after the tomatoes were harvested.
The tomatoes were then placed under controlled conditions, of 100% relative humidity, to facilitate fungal growth.
After 3 days, individual sepals were rated by 3 experts in terms of fungal presence and severity.
A Bayesian network was designed to combine the ratings from the experts in order to gain a single, more confident, rating.
The images from the first day were manually annotated to generate masks of individual sepals.
Manual labelling is tedious and prone to errors, especially at the sepal edges and the tip, and the manual masks still contain extra spectral pixels from the tomato skin.
To clean these masks an SVM was trained to segment the tomato crowns.
The cleaned masks are used for extracting the mean spectra of individual sepals which become their representative features.
Standard normal variate transformation is applied to the mean spectra in order to suppress scattering effects.
A Random Forest classifier was trained using repeated stratified 5-fold cross validation.
The independent test set is 30% of the overall data set.
Sepals showing only marginal, possibly ambiguous, signs of infection were eliminated from all data sets to improve overall correct assignment of ground truth labels.
The model trained on the training set reported 88.5% balanced accuracy on the validation set.
Testing on the independent set yielded a balanced accuracy of 77.4%. These results suggest NIR imaging may be suitable for automatically predicting the susceptibility of sepals of recently harvested tomatoes to future fungal infections.
Near-infrared (NIR) spectral imaging (1000-1700 nm) is explored for identification of the susceptible tomato sepals before any visual signs of infections appear.
Spectral images are acquired for the tomato cultivar Cappricia, a day after the tomatoes were harvested.
The tomatoes were then placed under controlled conditions, of 100% relative humidity, to facilitate fungal growth.
After 3 days, individual sepals were rated by 3 experts in terms of fungal presence and severity.
A Bayesian network was designed to combine the ratings from the experts in order to gain a single, more confident, rating.
The images from the first day were manually annotated to generate masks of individual sepals.
Manual labelling is tedious and prone to errors, especially at the sepal edges and the tip, and the manual masks still contain extra spectral pixels from the tomato skin.
To clean these masks an SVM was trained to segment the tomato crowns.
The cleaned masks are used for extracting the mean spectra of individual sepals which become their representative features.
Standard normal variate transformation is applied to the mean spectra in order to suppress scattering effects.
A Random Forest classifier was trained using repeated stratified 5-fold cross validation.
The independent test set is 30% of the overall data set.
Sepals showing only marginal, possibly ambiguous, signs of infection were eliminated from all data sets to improve overall correct assignment of ground truth labels.
The model trained on the training set reported 88.5% balanced accuracy on the validation set.
Testing on the independent set yielded a balanced accuracy of 77.4%. These results suggest NIR imaging may be suitable for automatically predicting the susceptibility of sepals of recently harvested tomatoes to future fungal infections.
Authors
H.A.C. de Villiers, A. Chauhan, E. Hogeveen-van Echtelt, M. Mensink, Z. Grboviæ, D. Stefanoviæ, M. Paniæ, S. Brdar
Keywords
fungus, NIR, hyperspectral, ‘Cappricia’, random forest
Online Articles (82)