Yield estimation and mapping in orchards are important for growers as it facilitates efficient utilization of resources and improves returns per unit area and time. With accurate knowledge of yield distribution and quantity, a grower can efficiently manage processes such as use of chemicals, fertilization, and thinning. Yield estimation also allows the grower to plan ahead of time their harvest logistics, crop storage, and sales (Aggelopoulou et al., 2011; Nuske, Wilshusen, Achar, Yoder, & Singh, 2014; Payne & Walsh, 2014). The standard approach to get yield information is currently manual sampling, which is labor-intensive, expensive, and often destructive (Gemtos, Fountas, Tagarakis, & Liakos, 2013). Constrained by these costs, sampling is often done over a few individual crops, and the measures are extrapolated over the entire farm. Inherent human sampling bias and sparsity in the measurements can result in inaccurate yield estimation.
Image annotation is the foundation behind many Artificial Intelligence (AI) products someone interacts with and is one of the most essential processes in Computer Vision (CV). In image annotation, annotators use tags, or metadata, to identify characteristics of the data that the end user requires from the AI model to learn to recognize. These tagged images are then used to train the computer to identify those characteristics when presented fresh, unlabeled data.
Also, defining the maturity level is of the highest importance because it makes the harvest as accurate as possible through the best classification of the crops and also supports the producer with the most effective resources management. With image annotation, any crop can be annotated in different labels, according to the maturity stage. As shown in the image below, the strawberry plant contains three maturity levels of fruits. Specifically, flowers are annotated with yellow bounding boxes and immature strawberries with the green ones. As for the mature strawberries, are annotated with red bounding boxes indicating that they are mature enough and ready for harvest.
Uses for machine vision detection of fruit in images of tree canopies include estimation of fruit number per tree (load), in-field fruit sizing, and automated harvest. Estimation of fruit size together with fruit number allows estimation of fruit weight (yield) per orchard. Fruit weight can be correlated to fruit lineal dimensions in many fruits. Such a relationship allows fruit weight to be estimated using machine vision, given a measure of the camera to fruit distance, e.g. using a time-of-flight (ToF) camera. For whole canopy images, as used in fruit load estimation, only a fraction of fruit in the image have entire outlines (e.g., not partially occluded), but that is sufficient to provide a size class distribution, assuming visible, outer canopy fruit are representative of all fruit on a tree (as achieved by Wang et al., 2017). Of course, a tree fruit load estimate relies on the assessment of the total number of fruit per tree, not the number of fruit visible in an image. With high accuracies reported for fruit detection in an image using deep learning methods, research attention should now shift to approaches to estimate total fruit per tree and per orchard.