Enhancing Green Fraction Estimation in Rice and Wheat Crops: A Self-Supervised Deep Learning Semantic Segmentation...

Area of Science:

• Agricultural Science
• Computer Vision
• Deep Learning

Background:

• Green fraction (GF) is crucial for assessing crop light interception and identifying high radiation use efficiency genotypes.
• Accurate GF estimation relies on high-quality segmentation datasets and robust image segmentation methods.
• Current annotation methods are costly and can lack accuracy, hindering GF monitoring.

Purpose of the Study:

• To develop a self-supervised strategy for deep learning semantic segmentation of rice and wheat field images.
• To enhance segmentation accuracy while reducing annotation costs for GF estimation.
• To bridge the reality gap between simulated and real-world field images.

Main Methods:

• Generated a large simulated dataset (sim dataset) of rice and wheat fields using the Digital Plant Phenotyping Platform.
• Employed CycleGAN for domain adaptation to create simulation-to-reality images (sim2real dataset).
• Trained and evaluated three semantic segmentation models (U-Net, DeepLabV3+, SegFormer) using real, sim, and sim2real datasets.

Main Results:

• SegFormer trained on the sim2real dataset achieved the highest segmentation accuracy for both rice (Accuracy=0.940, F1=0.937) and wheat (Accuracy=0.952, F1=0.935).
• This optimal strategy also yielded favorable GF estimation results (rice R²=0.967, wheat R²=0.984).
• The sim2real approach showed greater superiority for wheat than rice, potentially due to background differences.

Conclusions:

• The proposed self-supervised strategy effectively addresses high costs and annotation uncertainties in dataset creation.
• This method significantly enhances GF estimation accuracy for rice and wheat field images.
• The developed strategy offers a valuable tool for crop breeding and precision agriculture.