Search

[Display omitted] • Practitioner guidelines for the number of fruit to label for deep learning. • Predict AP score prior to training by calculating a novel similarity score. • Study of relationships between data-centric attributes and performance. • Evaluation on open-source datasets and a modern deep learning method (YOLOv5). Due to the rapid development of deep learning, object detection models have become the current tool of choice for on-tree fruit detection in precision agriculture. The pipeline of fruit detection based on deep learning generally starts from custom dataset collection, then image annotation, then training the object detection model, and finally determination of its accuracy and running the trained model for applications. To achieve better performance in fruit detection, most research has been focused on the third part of the pipeline which is improving or adjusting the state-of-art object detection models. However, the first two data-centric parts of the pipeline also require more investigation. For example, there is very limited research about how many annotations are sufficient for training and the degree of influence of image quality on the training performance for single-class fruit detection. Therefore, in this study, we thoroughly analysed seven public on-tree fruit datasets that cover apples, almonds, mangoes, and grape bunches under different image conditions. Our experiment for testing the size of the training dataset indicates that 2500 annotated objects are generally sufficient for single-class fruit training and our experiment for testing the object size shows simply that objects of a larger size have the potential to achieve better accuracy. Then a novel similarity score was proposed to allow the readers to easily estimate the expected Average Precision (AP) score without doing any training. The last two data-centric experiments then indicate that the influence of blurriness on the training accuracy is minor whereas less complex objects show the possibility of achieving better accuracy. Overall, such numerical data-centric analysis of on-tree fruit detection will enable us to better understand the influence on the training accuracy from data-centric attributes, which is of great benefit in helping practitioners prepare better quality datasets from a data-centric perspective and thus achieve higher training accuracy. [ABSTRACT FROM AUTHOR]

[Display omitted] • A novel method of estimating apple flower phenology distribution from images. • Algorithm validation on two varieties in both day and night-time orchard datasets. • Distributions were generated and compared at block, row and sub-tree levels. • Improved accuracy and speed when compared with YOLOv5. • The method was trained and evaluated with real field counts on common varieties. Estimation of phenology distribution in horticultural crops is very important as it governs the timing of chemical thinning in order to produce good quality fruit. This paper presents a novel phenology distribution estimation method named DeepPhenology for apple flowers based on CNNs using RGB images, which is able to efficiently map the flower distribution on an image-level, row-level, and block-level. The image classification model VGG-16 was directly trained with relative phenology distributions calculated from manual counts of flowers in the field and acquired imagery. The proposed method removes the need to label images, which overcomes difficulties in distinguishing overlapping flower clusters or identifying hidden flower clusters when using 2D imagery. DeepPhenology was tested on both daytime and night-time images captured using an RGB camera mounted on a ground vehicle in both Gala and Pink Lady varieties in an Australian orchard. An average Kullback-Leibler (KL) divergence value of 0.23 over all validation sets and an average KL value of 0.27 over all test sets was achieved. Further evaluation has been done by comparing the proposed model with YOLOv5 and shown to outperform this state-of-the-art object detection model for this task. By combining relative phenology distributions from a single image to a row-level or block-level distribution, we are able to give farmers a precise and high-level overview of block performance to form the basis for decisions on chemical thinning applications. [ABSTRACT FROM AUTHOR]

• A novel end-to-end flower segmentation algorithm for small and complex objects. • Algorithm evaluation on night and day time datasets in a commercial apple orchard. • A novel pipeline for autonomous side-view apple flower density mapping. Apple trees commonly require the removal of excessive flowers by thinning to produce high quality fruit. Machine vision has recently been applied to detect the flower density as the first step in this process. Existing work relying on color thresholding is sensitive to imaging conditions and the most recent published work using deep learning in this context has proven to be exceptionally slow to process. This paper presents an apple flower segmentation method on a pixel level based on a Fully Convolutional Network (FCN) together with a process of generating a map that can be used for a variable rate chemical sprayer. Despite the challenging conditions of an uncontrolled environment, our apple flower detector was able to generate a F 1 score at pixel-level up to 85.6%, which is a relatively high accuracy in terms of pixel-level segmentation. Our method has been tested on both daytime and night-time datasets, which strongly validates the ability of our apple flower detector to work under different conditions. The resulting detections are georeferenced and merged into a density map in the format necessary for application by a variable rate chemical sprayer. Finally, this flower density mapping system will benefit farmers by visualising the whole crop and extracting useful information to support their decision making for chemical thinning. [ABSTRACT FROM AUTHOR]

Counting grapevine shoots early in the growing season is critical for adjusting management practices but is challenging to automate due to a range of environmental factors. This paper proposes a completely automatic system for grapevine yield estimation, comprised of robust shoot detection and yield estimation based on shoot counts produced from videos. Experiments were conducted on four vine blocks across two cultivars and trellis systems over two seasons. A novel shoot detection framework is presented, including image processing, feature extraction, unsupervised feature selection and unsupervised learning as a final classification step. Then a procedure for converting shoot counts from videos to yield estimates is introduced. The shoot detection framework accuracy was calculated to be 86.83% with an F1-score of 0.90 across the four experimental blocks. This was shown to be robust in a range of lighting conditions in a commercial vineyard. The absolute predicted yield estimation error of the system when applied to four blocks over two consecutive years ranged from 1.18% to 36.02% when the videos were filmed around E-L stage 9. The developed system has an advantage over traditional PCD mapping techniques in that yield variation maps can be obtained earlier in the season, thereby allowing farmers to adjust their management practices for improved outputs. The unsupervised feature selection algorithm combined with unsupervised learning removed the requirement for any prior training or labeling, greatly enhancing the applicability of the overall framework and allows full automation of shoot mapping on a large scale in vineyards. [ABSTRACT FROM AUTHOR]