Deep learning-based semantic segmentation for rice yield estimation by analyzing the dynamic change of panicle...

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Phenotyping

Background:

• Increasing global population and climate change demand higher agricultural productivity.
• Existing rice yield prediction models often lack dynamic temporal analysis, limiting accuracy.
• Dynamic changes in rice panicle coverage significantly impact yield, necessitating advanced monitoring techniques.

Purpose of the Study:

• To develop a novel temporal framework for rice phenotyping and yield prediction.
• To integrate high-resolution RGB imagery with deep learning for dynamic panicle analysis.
• To accurately estimate rice yield and its components using time-series imaging data.

Main Methods:

• Acquired high-resolution RGB images of rice canopies over two growing seasons.
• Evaluated five semantic segmentation models (DeepLabv3+, U-Net, PSPNet, FPN, LinkNet) for rice panicle delineation.
• Fitted time-series panicle coverage data to a piecewise function to extract key growth parameters (K, g, d0, a, d1) and used these in machine learning models (PLSR, RFR, GBR, XGBR) for yield prediction.

Main Results:

• DeepLabv3+ and LinkNet achieved high performance in panicle segmentation (mIoU > 0.81).
• Maximum panicle coverage (K) strongly correlated with Yield (r=0.87) and Grain Number (r=0.85).
• Random Forest Regressor (RFR) and XGBoost Regressor (XGBR) models showed the highest yield prediction accuracy (R²=0.89).

Conclusions:

• The developed framework effectively quantifies rice developmental dynamics using RGB imagery.
• This approach enables accurate yield prediction, supporting precision agriculture and crop breeding.
• Integrating temporal imaging with deep learning offers a powerful tool for agricultural monitoring.