Automatic flower cluster estimation in apple orchards using aerial and ground based point clouds

Chemical and mechanical thinning processes have long been used in stone and pome fruit production. During the thinning of apple flowers, growers use chemicals to regulate the tree load. Hand thinning is applied after the June drop to prune trees with excess crop load. The process of thinning can be unpredictable especially in biennial bearing cultivars. Thus, incentives to optimise chemical usage and to reduce expensive manual labour is ever increasing. Ground based machine vision systems have grown in popularity in orchard management due to the level of detail as well as plant coverage they can inspect with. Additionally, unmanned aerial vehicles (UAV)-based remote sensing technology is becoming a popular non-invasive quality inspection solution. This work proposes a framework for combining UAV and ground based RGB image data to detect flowering intensity in a Dutch Elstar apple orchard. The framework, based on point cloud reconstruction, presents automatic point cloud handling techniques as well as automated unsupervised flowering intensity estimation methods. Two linear regression models based on unsupervised machine learning methods were trained and validated from the framework that estimate flowering intensity in the orchard with both models having R-2 > 0.65, RRMSE < 20% and p-stat < 0.005 for the correlation between the image derived flower index and the flower cluster number counted in field. The proposed methods provide a novel strategy for guiding flower thinning using simple RGB images and location data only. Moreover, the proposed methods also reveal the flexibility of intra-tree inspection by checking its sub-volumes. (C) 2022 The Author(s). Published by Elsevier Ltd on behalf of IAgrE.

Automated summary

This study proposes a framework for combining UAV and ground based RGB image data to detect flowering intensity in a Dutch Elstar apple orchard. Two linear regression models based on unsupervised machine learning methods were trained and validated, achieving good results. The proposed methods can guide flower thinning and provide flexible intra-tree inspection.