??AI for Pears: Pollen Estimation Using YOLO for Optimized Collection

Pear pollination relies heavily on artificial methods due to the instability of insect pollination under adverse weather conditions.

Manual pollen collection is labour-intensive, and Japan increasingly depends on imported pollen, raising costs and risks of insufficient supply.

Efficiently estimating and collecting pear pollen domestically is essential to mitigate these issues.

Methodology

The study developed an AI-based solution using the YOLO (You Only Look Once) deep learning model for object detection. Key steps included:

1. Classification and Detection: YOLO classified pear flowers into five stages: immature, balloon-shaped (optimal), blooming, blooming end (too late for collection), and others.
2. Training and Validation: The model was trained on 1271 images of pear flowers and validated using 254 images. The training included 400 epochs with a batch size of 16, using YOLOv8n for efficient detection.
3. Estimation: Flower detection results were combined with experimentally derived pollen yields per flower at each stage to estimate total pollen amounts for branches.

In the figure above: ?(a) Loss rate for detection frame. (b) Loss rate for detection class. (c) mAP50.

In the figure above: (a) Late, Too late and Other mixed. (b) Early and Best mixed. (c) Early

Key Discoveries

The study revealed that YOLO-based flower detection achieved a strong overall performance, with a mean Average Precision (mAP50) of 76% at 330 epochs. Classification accuracy for the flowering stages varied significantly, with early and blooming stages achieving 80% and 86% accuracy, respectively.

However, the balloon-shaped "best" stage, critical for pollen collection, had a lower accuracy of 69%, mainly due to visual similarities with the early stage. The "other" category had the lowest accuracy (54%) because many flowers were undetectable in certain images.

Pollen yield analysis demonstrated that the balloon-shaped stage consistently produced the highest amount of pollen across pear varieties. For example, Shinko pears yielded up to 1.347 mg per flower in this stage, while Chojuro and Nepal pears produced 0.757 mg and 1.137 mg, respectively.

Pollen yields declined drastically in later flowering stages, with almost no viable pollen available in the blooming end stage. These findings underscore the importance of accurate detection during the optimal balloon-shaped stage to maximize pollen collection.

The AI-based estimation of pollen amounts showed promise but also highlighted challenges.         

The estimated pollen amounts were consistently lower than actual measured values, with errors ranging from -18.02% to -37.7% across pear varieties. These discrepancies were attributed to undetected flowers hidden behind branches, misclassifications of flowering stages, and variability in measured pollen yields per flower.

Despite these challenges, the method's ability to estimate pollen yields in real time represents a significant advancement over traditional temperature-based prediction models, which lack such responsiveness.

In the figure above: (a) Chojuro. (b) Shinko. (c) Nepal. (d) Total of 3 varieties.

The potential for real-time estimation makes this technology particularly valuable for optimizing pollen collection. By tracking flowering stages daily, growers can identify the best time to maximize yields, thus enabling a practical application for smart agriculture. The study demonstrates a promising step forward in leveraging AI for efficient and cost-effective pear pollen collection.

Practical Importance

The proposed AI-based technology is a step towards automating and optimizing pear pollen collection, offering:

• Labor Efficiency: Reduces manual effort by automating flower classification and pollen yield estimation.
• Cost Reduction: Minimizes reliance on imported pollen, lowering expenses and ensuring a stable supply.
• Precision Agriculture: Allows growers to target the optimal collection period for maximum yield, improving productivity.

This study demonstrates the feasibility of using deep learning for efficient pear pollen estimation and highlights areas for further improvement, such as enhancing detection accuracy for hidden flowers and refining yield measurement techniques. The approach contributes to smart agricultural practices and could be expanded to other crops and applications.

Citation

Endo, K., Hiraguri, T., Kimura, T. et al. Estimation of the amount of pear pollen based on flowering stage detection using deep learning. Sci Rep 14, 13163 (2024). https://doi.org/10.1038/s41598-024-63611-w

??Direct Reference

?? What's on

If you'd like to receive the regular 'AI in Agriculture' newsletter in your inbox, simply add your email to my mailing list.

Join almost 9,600 readers who enjoy weekly updates on AI advancements in agriculture!

AI for Plant Phenotyping and Quality Assurance

To get more information about plant phenotyping capabilities and grapevine phenotyping capabilities with mobile - ask Petiole Pro