Deep Learning Methods Using Imagery from a Smartphone for Recognizing Sorghum Panicles and Counting Grains at a Plant...

Area of Science:

• Agricultural Science
• Computer Science
• Plant Biology

Background:

• High-throughput phenotyping is crucial for crop improvement but faces bottlenecks in field trait characterization.
• Accurate yield estimation in sorghum (Sorghum bicolor L.) relies heavily on quantifying grain number per panicle.
• Integrating computer vision and artificial intelligence (AI) offers a solution to reduce labor and time in field phenotyping.

Purpose of the Study:

• To enhance sorghum panicle detection and grain number estimation using smartphone-captured images.
• To develop and validate AI models for accurate, field-based yield prediction in sorghum.
• To establish a foundation for a robust, plant-level sorghum yield estimation application.

Main Methods:

• Collected and manually labeled a benchmark dataset of 648 smartphone images of sorghum panicles.
• Trained detection and segmentation models (Detectron2, Yolov8) for panicle identification.
• Developed and trained three grain counting models (MCNN, TCNN-Seed, Sorghum-Net), with Sorghum-Net being novel.

Main Results:

• Yolov8 achieved 89% average precision for panicle detection and segmentation.
• The developed Sorghum-Net model demonstrated a 17% mean absolute percentage error for grain number estimation.
• A simple equation was derived to correlate model counts with field-observed grain numbers, achieving 17% overall error.

Conclusions:

• The proposed AI framework effectively estimates sorghum grain number from smartphone images with high accuracy.
• This approach significantly reduces labor and time compared to traditional phenotyping methods.
• The study provides a scalable foundation for developing advanced applications for real-time sorghum yield estimation.