Recent years have seen a remarkable growth trend in avocado consumption, with projections that its exports will overpass those of pineapple by 2030, becoming the second most traded tropical fruit, just surpassed by the banana. When paired with their high unit value, this growth could make the avocado one of the most important fruit commodities of the next decades. The economic relevance of avocado production has influenced its expansion into new regions of the world. Portugal is now the second largest producer of avocados in Europe, taking advantage of the climatic adequacy of the Algarve region to the requirements of this produce. As the production of avocados is still limited to the tropical and subtropical regions, its exports are directly impacted by time-consuming distribution channels. Combined with the relatively high unpredictability of their post-harvest behaviour, this makes avocado fruits highly prone to wasteful practices. The development of non-destructive tools that accurately trace the ripening process of avocado fruits could then be key to a better management of their shelf-life, optimizing their post-harvest handling to a point of drastically reducing distribution waste. This project aimed to develop a smart data-driven tool that uses Machine Learning to improve the traceability of the ripening process of Hass avocado pears. A total of 476 avocados were divided between three storage groups, with different environmental conditions, and their ripening behaviour was traced by the implementation of an innovative 5-stage Ripening Index, that classified the ripening stage of each sample, daily, according to a set of common traits. This information was paired with daily photographs of each avocado, to build a database of labelled image data that was then fed to two Convolutional Neural Networks, AlexNet and ResNet-18, taking advantage of the concept of transfer-learning where pre-trained knowledge is used to improve their adaptation to new sets of data. The networks were trained to recognise the specific visual traits of each ripening stage, so that they could predict the state of new unlabelled data. This knowledge was tested on new datasets, reaching an average final accuracy of 77,8%, with an average of 88,4% of the predictions falling within a half-stage margin of error, and an average of 95,0% within one stage of the attributed classifications. Both pre-trained networks achieved similar performances, with a slight advantage for ResNet-18, and the results were also similar across storage groups, suggesting that the predictive accuracy wasn’t affected by the handling of the fruits. These results represent an important step for the integration of Computer Vision tools on the post-harvest management of perishable products, which could not only improve shelf-life determinations, but ultimately be expanded into other assessments, with a major potential impact on waste prevention and quality improvement.
- Avocado
- Smart data
- Ripening
- Shelf-life
- Data-driven prediction
- Mestrado em Engenharia Alimentar
Smart data-driven tool for predicting avocado fruits (persea americana) shelf-life
Xavier, P. A. E. L. (Student). 10 Jan 2023
Student thesis: Master's Thesis