Estimating the Ripeness of Hass Avocado Fruit Using Deep Learning with Hyperspectral Imaging

Abstract

Rapid ripeness assessment of fruit after harvest is important to reduce post-harvest losses by sorting fruit according to the duration until they become ready to eat. However, there has been little research on non-destructive estimation of the ripeness and ripening speed of avocado fruit. Unlike previous methods, which classify the ripeness of fruit into a few categories (e.g., unripe and ripe) or indirectly estimate ripeness from its firmness, we developed a method using hyperspectral imaging coupled with deep learning regression to directly estimate the duration until ripeness of Hass avocado fruit. A set of 44,096 sub-images of 551 Hass avocado fruit images was used to train, validate, and test a convolutional neural network (CNN) to predict the number of days until ripeness. Training, validation, and test samples were generated as sub-images of Hass fruit images and were used to train a spectral–spatial residual network to estimate the duration to ripeness. We achieved predictions of duration to ripeness with an average error of 1.17 days per fruit on the test set. A series of experiments demonstrated that our deep learning regression approach outperformed classification approaches that rely on dimensionality reduction techniques such as principal component analysis. Our results show the potential for combining hyperspectral imaging with deep learning to estimate the ripeness stage of fruit, which could help to fine-tune avocado fruit sorting and processing.