Image 6_Deep learning-based semantic segmentation for rice yield estimation by analyzing the dynamic change of panicle coverage.tif

Introduction<p>Rising global populations and climate change necessitate increased agricultural productivity. Most studies on rice panicle detection using imaging technologies rely on single-time-point analyses, failing to capture the dynamic changes in panicle coverage and their effects on yie...

Full description

Saved in:
Bibliographic Details
Main Author: Hyeok-Jin Bak (22051985) (author)
Other Authors: Eun-Ji Kim (257634) (author), Ji-Hyeon Lee (254867) (author), Sungyul Chang (579208) (author), Dongwon Kwon (22051988) (author), Woo-Jin Im (22051991) (author), Woon-Ha Hwang (22051994) (author), Jae-Ki Chang (22051997) (author), Nam-Jin Chung (22052000) (author), Wan-Gyu Sang (9547140) (author)
Published: 2025
Subjects:
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Introduction<p>Rising global populations and climate change necessitate increased agricultural productivity. Most studies on rice panicle detection using imaging technologies rely on single-time-point analyses, failing to capture the dynamic changes in panicle coverage and their effects on yield. Therefore, this study presents a novel temporal framework for rice phenotyping and yield prediction by integrating high-resolution RGB imagery with deep learning-based semantic segmentation.</p>Methods<p>High-resolution RGB images of rice canopies were acquired over two growing seasons. We evaluated five semantic segmentation models (DeepLabv3+, U-Net, PSPNet, FPN, LinkNet) to effectively delineate rice panicles. Time-series panicle coverage data, extracted from the segmented images, were fitted to a piecewise function to model their growth and decline dynamics. This process distilled key predictive parameters: K (maximum panicle coverage), g (growth rate), d0 (time of maximum growth rate), a (decline rate), and d1 (transition point). These parameters served as predictors in four machine learning regression models (PLSR, RFR, GBR, and XGBR) to estimate yield and its components.</p>Results<p>In panicle segmentation, DeepLabv3+ and LinkNet achieved superior performance (mIoU > 0.81). Among the piecewise function parameters, K showed the strongest positive correlation with Yield and Grain Number (GN) (r = 0.87 and r = 0.85, respectively), while d0 was strongly negatively correlated with the Filled Grain Ratio (FGR) (r = -0.71). For yield prediction, the RFR and XGBR models demonstrated the highest performance (R<sub>2</sub>= 0.89). SHAP analysis quantified the relative importance of each parameter for predicting yield components.</p>Discussion<p>This framework proves to be a powerful tool for quantifying rice developmental dynamics and accurately predicting yield using readily available RGB imagery. It holds significant potential for advancing both precision agriculture and crop breeding efforts.</p>