PanicleNeRF: Low-Cost, High-Precision In-Field Phenotyping of Rice Panicles with Smartphone

Area of Science:

• Agricultural Science
• Computer Vision
• Genetics

Background:

• Rice panicle traits are crucial for grain yield and a key focus in phenotyping.
• Existing phenotyping methods often require controlled indoor settings, limiting field applicability.
• Accurate in-field measurement of rice panicle traits is essential for breeding programs.

Purpose of the Study:

• To develop a novel, low-cost, high-precision method for 3D rice panicle reconstruction in natural field conditions.
• To enable high-throughput phenotyping of rice panicles using smartphone-acquired videos.
• To facilitate accelerated breeding of improved rice varieties.

Main Methods:

• Developed PanicleNeRF, integrating Segment Anything Model (SAM) and YOLOv8 for precise 2D image segmentation.
• Utilized Neural Radiance Fields (NeRF) for 3D model reconstruction from segmented images.
• Processed 3D point clouds to extract key panicle traits and assess their correlation with yield components.

Main Results:

• PanicleNeRF achieved high segmentation accuracy (mean F1 score 86.9%, IoU 79.8%) and superior boundary overlap compared to YOLOv8.
• 3D reconstruction quality significantly surpassed traditional Structure-from-Motion Multi-View Stereo (SfM-MVS) methods.
• Accurate extraction of panicle length (rRMSE < 3%) and strong correlations between estimated volume and grain number (R² > 0.82) and mass (R² > 0.76).

Conclusions:

• PanicleNeRF offers an effective, low-cost solution for in-field rice panicle phenotyping.
• The method enables high-throughput data acquisition, accelerating rice breeding efficiency.
• This approach overcomes limitations of indoor-based phenotyping, providing valuable field data.