Estimation of Apple Fruitlet Density in Orchards using Deep Learning Techniques

As the global population increases, there is pressure to decrease the cost and reliance on human labour of farming. Deep learning image analysis for agricultural applications is a growing area of research, however most studies in the field of fruit detection have only focused on ripe fruit for mechanical harvesting. Commercial apple orchards require the thinning of fruitlets to maintain the quality of the crop, which is largely done by hand and is the second largest expense in commercial apple production. This paper proposes a deep learning method of assessing the density of fruitlets to enable chemical thinning, which requires the detection of fruit at an earlier developmental stage than has been previously researched. The novel proposed pipeline could be used to apply more accurate volumes of chemical thinner or to apply variable volumes based on the requirements of individual clusters that would significantly reduce the costs associated with hand-thinning. This study seeks to examine the suitability of four existing object detection architectures for detecting fruitlet clusters and four existing classification architectures for counting the number of fruitlets in each cluster. The detection networks achieved a maximum recall of 0.480 and the most successful classification network achieved at least 99% accuracy for each class.