Avocado stem-end rot detection using hyperspectral imaging

One of the most common postharvest diseases impacting avocado is stem-end rot, a fungal disease whose damage is usually not externally visible. Only after extensive damage do symptoms become visible to the naked eye. Currently infections are primarily detected using the destructive approach (cutting the fruit open). There is significant interest in identifying non-destructive approaches testing entire avocado batches for stem-end rot. We explore visible-near infrared, VIS/NIR (wavelengths 470-1000 nm), hyperspectral imaging and machine learning to non-destructively detect stem-end rot in ‘Hass’ avocados. The data set consisted of 480 avocados, 240 of which were untreated and 240 inoculated (with stem-end rot). An independent test set of 80 avocados was also collected. After pre-processing, spectra from avocado stem pixels (and neighbouring regions) is averaged to obtain mean spectral features. These features are then used for training the Random Forest, XGBoost and a semi-supervised deep autoencoder classifiers at the task of distinguishing between healthy and infected samples. Overall, all classifiers led to similar performance, obtaining between 75 and 83% accuracy over the validation set. However, testing on the independent test set yielded a significant drop in performance with accuracy in the range of 56-65%, as well as elevated levels of false positives, while false negatives remained low. The results suggest that VIS/NIR based models may not be directly applicable for predicting stem-end rot in avocados. We hypothesize that spectral data of independent samples is not sufficient to capture stem-end rot because maturity/age and linked physiological changes might play an important factor, which was not considered in the current study.