Deep Learning-Based Semantic Segmentation for Rice Yield Estimation by Analyzing the Dynamic Change of Panicle Coverage

Hyeok-Jin Bak¹Eun-Ji Kim¹Ji-Hyeon Lee¹Sungyul Chang¹Dongwon Kwon¹Woo-Jin Im¹Woon-Ha Hwang¹Jae-Ki Chang¹Nam-Jin Chung²Wan-Gyu Sang^1*

• ¹National Institute of Crop and Food Science, Rural Development Administration, Jeonbuk-do, Republic of Korea
• ²Jeonbuk National University, Department of Agronomy, Jeonbuk-do, Republic of Korea

Rising global populations and climate change necessitate efficient agricultural practices, demanding increased productivity. Most studies on rice panicle detection using imaging technologies focus on single-time-point analyses, failing to capture dynamic changes in panicle coverage and their effects on yield. Therefore, this study presents a novel temporal method for rice phenotyping and yield prediction, integrating high-resolution RGB imagery and deep learning-based semantic segmentation. Images of rice canopies were acquired with fixed-position and handheld cameras under field and soilbin conditions across two growing seasons. Five semantic segmentation models (DeepLabv3+, U-Net, PSPNet, FPN, LinkNet) with ResNet-50 and ResNet-101 backbones were evaluated for effectiveness in delineating rice panicles. DeepLabv3+ and LinkNet consistently achieved superior performance (mIoU > 0.81). Time-series panicle coverage data, extracted from segmented images, were fitted to a piecewise function to model sigmoidal growth and quadratic decline. Modeling distilled complex temporal dynamics of panicle development into key predictive parameters. These parameters -K (maximum panicle coverage), g (growth rate), d₀ (time of maximum growth rate), a (decline rate), and d₁ (transition point) served as predictors for four machine learning regression models (partial least squares regression, PLSR; random forest regressor, RFR; gradient boosting regressor, GBR; and XGBoost regressor, XGBR) to predict yield and yield components (grain number (GN), panicle number (PN), number of grains per panicle (GNP), 1000-grain weight (TGW), and filled grain ratio (FGR)). Additionally, the study evaluated effects of nitrogen fertilization (0, 98.8, and 197.6 kg ha⁻¹), transplantation date (regular and late), and rice variety (Nampyeong, Shindongjin, Dongjin-1, and Saeilmi) on panicle coverage dynamics and yield. K had the strongest positive correlation with yield and GN (r = 0.87 and r = 0.85, respectively), whereas d₀ was strongly negatively correlated with FGR (r = -0.71). Specifically for yield prediction, RFR and XGBR (R² = 0.89) achieved the highest performance, followed by GBR (R² = 0.88) and PLSR (R² = 0.82). SHAP analysis quantified the relative importance of piecewise function parameters for each yield component. This framework demonstrates potential for accurate rice phenotyping and yield prediction using readily available RGB imagery, providing a valuable tool for precision agriculture and crop improvement.Introduction