Assessment of Mixed Sward Using Context Sensitive Convolutional Neural Networks.

Area of Science:

• Agricultural Science
• Computer Science
• Remote Sensing

Background:

• Traditional methods for measuring forage biomass are labor-intensive and limit breeding efficiency.
• High-throughput phenotyping using automation and remote sensing offers a solution to bottlenecks in plant breeding.

Purpose of the Study:

• To develop and evaluate a deep learning model for accurate biomass estimation in mixed white clover and perennial ryegrass swards.
• To address the challenge of high occlusion in dense pasture canopies using advanced image analysis.

Main Methods:

• Utilized RGB imaging and a novel convolutional neural network (CNN) architecture named the local context network (LC-Net).
• LC-Net was designed for semantic segmentation of dense pasture, specifically differentiating white clover and perennial ryegrass.
• Evaluated model performance using metrics such as mean accuracy and mean intersection over union on a dedicated dataset.

Main Results:

• The LC-Net achieved a mean accuracy of 95.4% and a mean intersection over union of 81.3% for segmenting clover from ryegrass.
• The developed model outperformed existing literature methods for semantic segmentation in mixed swards.
• Despite high segmentation accuracy, direct correlation with harvested dry-matter showed limited improvement, indicating a need for further refinement.

Conclusions:

• The LC-Net demonstrates significant potential for automated biomass estimation in forage crops through precise image segmentation.
• Combining RGB data with complementary sensor information (e.g., volumetric data) may enhance biomass estimation accuracy in future research.
• This automated approach facilitates high-throughput phenotyping, accelerating breeding programs for higher-yielding forage species.