Robust soybean seed yield estimation using high-throughput ground robot videos.

Area of Science:

• Agricultural Science
• Computer Vision
• Deep Learning

Background:

• Traditional soybean yield data collection is labor-intensive, costly, and prone to equipment failure.
• Computer vision offers a potential solution for extracting detailed yield information directly from images.

Purpose of the Study:

• To present a novel, efficient method for soybean yield estimation using computer vision and deep learning.
• To develop a scalable solution for agricultural breeding programs and productivity enhancement.

Main Methods:

• Utilized a ground robot with fisheye cameras to capture soybean plot videos.
• Developed and applied the P2PNet-Yield deep learning model for seed counting and yield regression.
• Incorporated fisheye image correction and data augmentation for improved accuracy and generalizability.

Main Results:

• The P2PNet-Yield model achieved a genotype ranking accuracy score of up to 83%.
• Demonstrated a reduction of up to 32% in time and associated costs for yield data collection.
• Validated the model using two years of yield testing plot data (8,500 plots in 2021, 650 in 2023).

Conclusions:

• The novel computer vision and deep learning approach provides a more efficient and accurate method for soybean yield estimation.
• This technology offers a scalable solution for enhancing agricultural breeding programs and overall productivity.