Viability of avocado yield quantification using machine vision and machine learning : a thesis presented in partial fulfilment of the requirements for the degree of Master of Engineering at Massey University, Manawatū, New Zealand

Abstract

The avocado is the green celebrity of the fruit world with NZ exports more than doubling since 2006. Quantitative crop estimation is a critical piece of information in a horticultural supply chain, as it heavily dictates planning of harvest timing, labour, and marketing. Like all fruit crops, the avocado industry has a need to estimate their harvest volume each season. And so arises the question many growers ask; how can they accurately estimate their harvest? However, factors like the close colour matching of the fruit with the canopy, and the large canopy of the production trees, create some challenges for crop estimation. This work provides an insight into the current methods that have been explored and what direction growers should move towards, to increase their ability to estimate crop load. Research solutions options were rated using a weighting matrix that considers accuracy, availability, cost, and relevance in order to justify technologies that are most suited to further development. A novel labelled avocado dataset was used to train a convolutional neural network called YOLOv5. The trained YOLOv5 model showed promising results, with a real time detection accuracy of 88%, and a model mAP of 86.9% at 0.6 IoU threshold when trained at an image size of 512x512 pixels. YOLOv5 predicted fruit per tree counts were within 5% compared to the harvest fruit per tree count.