Wheat-Net: An Automatic Dense Wheat Spike Segmentation Method Based on an Optimized Hybrid Task Cascade Model.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate wheat spike segmentation is crucial for estimating yield and analyzing spike morphology.
• Existing methods struggle with complex backgrounds and dense, occluded wheat spikes in field conditions.

Purpose of the Study:

• To develop and validate a novel instance segmentation method for precise wheat spike detection and counting.
• To improve the accuracy of image-based wheat phenotyping in challenging field environments.

Main Methods:

• Utilized a Hybrid Task Cascade model with Res2Net50 as the backbone.
• Incorporated multi-scale training, deformable convolutional networks, and Generic ROI Extractor for enhanced feature learning.
• Trained and validated the model on diverse field-based wheat images with varying environmental conditions.

Main Results:

• Achieved high average precision (AP) of 0.904 for bounding box and 0.907 for mask segmentation.
• Demonstrated exceptional wheat spike counting accuracy of 99.29%.
• Wheat-Net showed robust performance on challenging datasets with dense, occluded spikes and varied backgrounds.

Conclusions:

• The proposed Wheat-Net method effectively addresses dense wheat spike detection and segmentation challenges.
• This technology offers significant potential for improving wheat yield estimation and spike morphology assessments.
• The model's performance validates its utility for automated, image-based agricultural phenotyping.