?? AI for Dragon Fruit: Ripeness of Pitaya in Natural Environments ??

Pitaya, commonly known as dragon fruit, is a tropical fruit originating from the cactus species of the genus Hylocereus and Selenicereus.

It is widely cultivated in regions with warm climates, particularly in southern China, Southeast Asia, Central and South America. The fruit is known for its vibrant appearance, with an outer skin that ranges from bright red to yellow, adorned with distinctive green scales.

Inside, it contains flesh that is either white, red, or purple, dotted with small, edible black seeds. Pitaya is not only prized for its striking look but also for its nutritional benefits, as it is rich in antioxidants, vitamins, and fiber.

In the figure above: (A-H) Sequential illustrations progression of reproductive development, flowering, fruit development and maturation of dragon fruit (H. undatus) according to the extended BBCH scale from Kishore et al. (2016). Time elapsed in each stage (horizontal bar), (2018-2020). Reproductive development (A, B); Flowering (C, D); Fruit development (E, F); Fruit maturity (G, H)

Due to its growing popularity in global markets and its significant role in export industries, especially in China, efficient and timely harvesting of pitaya is crucial for minimizing waste and optimizing production.

AI for Pitaya

The research paper presents a comprehensive study on developing an AI-based model for identifying the ripeness of pitaya (dragon fruit) to advance precision agriculture in China.

The focus of the study is on addressing the inefficiencies of manual fruit harvesting and improving the detection and classification of pitaya ripeness using machine vision technology.

The authors propose an improved model based on YOLOv8n, named GSE-YOLO, which integrates multiple improvements for better detection accuracy, computational efficiency, and robustness.

Key Objectives

1. Improve accuracy in identifying pitaya ripeness across varying natural conditions, where factors like light, overlapping fruits, and environmental complexity can challenge detection models.
2. Enhance computational efficiency by reducing the model’s size and computational complexity while maintaining high detection accuracy.

Research Problem

Pitaya production in China is labor-intensive, inefficient, and costly. The conventional methods of determining ripeness are highly manual, leading to wastage and increased production costs. The aim is to automate this process using a lightweight, precise AI model that can accurately detect the ripeness of pitaya across different stages.        

The GSE-YOLO Model

GSE-YOLO is a modified version of the YOLOv8n object detection algorithm, specifically designed to classify pitaya fruits into four ripeness stages: Bud, Immature, Semi-mature, and Mature.

Key Modifications to YOLOv8n

GhostConv: Replaces the standard convolutional layers in YOLOv8n to reduce computational parameters while maintaining accuracy.

1. SPPELAN Pyramid Pooling Structure: Enhances feature extraction by replacing the traditional SPPF with SPPELAN, which improves speed and accuracy by efficiently extracting key features.
2. EMA-Attention Mechanism: Improves the model’s attention to important features while reducing the computational burden.
3. WIoU Loss Function: Modifies the original CIoU loss function to improve the model’s convergence, robustness, and generalization ability.

Data and Experiment Setup

• Dataset: A total of 2,788 images of pitaya at different ripeness stages were used. The data was divided into a training set (70%), validation set (20%), and test set (10%).
• Computational Setup: The experiments were conducted on a system with an Intel Core i9-13900K CPU and an NVIDIA GeForce RTX 4090 GPU, using the PyTorch framework.

Results

The improved GSE-YOLO model achieved significant performance gains across multiple metrics:

• Detection Accuracy: 85.2%
• Recall Rate: 87.3%
• F1-Score: 86.23
• mAP50 (mean Average Precision at 50% Intersection over Union): 90.9%

Breakdown of Results by Ripeness Stage

• Bud: Precision of 97.1%, recall of 84.2%, F1-score of 90.19%, and mAP50 of 92.3%.
• Immature: Precision of 94.2%, recall of 82.2%, F1-score of 87.79%, and mAP50 of 90.4%.
• Semi-mature: Precision of 54.9%, recall of 93.1%, F1-score of 69.07%, and mAP50 of 84.8%.
• Mature: Precision of 94.6%, recall of 89.7%, F1-score of 92.08%, and mAP50 of 96.2%.

Comparison with Other Models

The GSE-YOLO model was compared to other state-of-the-art object detection models like YOLOv5n, YOLOv6n, YOLOv7, and the original YOLOv8n. In terms of average precision (mAP50), the improved GSE-YOLO model outperformed these models:

• YOLOv5n: mAP50 of 89.4%
• YOLOv6n: mAP50 of 87.7%
• YOLOv7: mAP50 of 88.3%
• YOLOv8n (baseline): mAP50 of 89.7%
• GSE-YOLO (improved): mAP50 of 90.9%

Efficiency and Model Size

• GSE-YOLO has a model size of only 5.30 MB, which is smaller than YOLOv7 (71.3 MB) and comparable to YOLOv5n (3.64 MB) but with higher accuracy.
• The number of parameters in the GSE-YOLO model is 2,660,358, which is significantly lower than YOLOv7’s 37,212,738, making it more suitable for deployment in resource-constrained environments.

Conclusion

The GSE-YOLO model demonstrates a high level of precision in detecting the ripeness of pitaya, with accuracy metrics close to 90%. It successfully addresses the challenges of complex environments, such as lighting and overlapping fruits, by incorporating various improvements in the YOLOv8n architecture. The lightweight nature of the model, combined with its robust performance, makes it an excellent candidate for real-world applications in precision agriculture, helping to reduce labor costs and improve harvesting efficiency in the pitaya industry.

Future Implications

The research highlights the potential of AI and machine learning models in advancing precision agriculture. The GSE-YOLO model is a step towards automating fruit detection and harvesting, reducing wastage, and improving productivity in agricultural practices. Further enhancements in the model could include adaptations for different fruit types and further optimization for even greater detection speeds.

Key Numerical Insights

• Detection accuracy: 85.2%
• Recall rate: 87.3%
• F1-Score: 86.23
• mAP50: 90.9%
• Model size: 5.30 MB
• Number of parameters: 2,660,358
• Precision by ripeness stage: Bud: 97.1% Immature: 94.2% Semi-mature: 54.9% Mature: 94.6%

Citation

Qiu, Z.; Huang, Z.; Mo, D.; Tian, X.; Tian, X. GSE-YOLO: A Lightweight and High-Precision Model for Identifying the Ripeness of Pitaya (Dragon Fruit) Based on the YOLOv8n Improvement. Horticulturae 2024, 10, 852. https://doi.org/10.3390/horticulturae10080852

AI in Agriculture Podcast: Exploring Pitaya Ripeness

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 over 9,000 readers who enjoy weekly updates on AI advancements in agriculture!

AI for Plant Phenotyping and Quality Assurance

To get more information about crop phenotyping capabilities with mobile - ask Petiole Pro