Rice tiller number estimation based on an improved Swin-UNet model and multi-feature fusion

Area of Science:

• Agricultural Science
• Plant Breeding
• Computer Vision

Background:

• Rice tillering characteristics are crucial for high-yield breeding, requiring precise monitoring.
• Drone imagery offers high-throughput monitoring but faces challenges in complex environments and dense plots.
• Accurate tiller number estimation is vital for smart breeding and germplasm screening.

Purpose of the Study:

• To develop an advanced method for high-throughput, accurate rice tiller number estimation using drone imagery.
• To improve rice plant segmentation and tiller counting in challenging early tillering stages.
• To provide quantitative data for tillering trait identification in rice breeding programs.

Main Methods:

• An improved Swin-UNet model was utilized for enhanced rice plant segmentation.
• Multi-feature fusion combined morphological and color data for tiller estimation.
• A Particle Swarm Optimization (PSO)-optimized XGBoost model was employed for tiller number prediction.

Main Results:

• The improved Swin-UNet model achieved 92.5% segmentation accuracy, a 7.2% improvement over U-Net.
• The PSO-XGBoost model, using 12 selected features, demonstrated strong performance with R²=0.85 and RMSE=0.35.
• Tiller number thematic maps were successfully generated for 576 breeding plots.

Conclusions:

• The proposed method effectively addresses challenges in early-stage rice tiller monitoring.
• This approach provides robust data support for germplasm tillering trait identification.
• The study significantly advances smart breeding by enabling precise, high-throughput phenotyping.