Deep Learning-Based Methods for Phenotypic Trait Extraction in Rice Panicles

Zhiao Wang^1*Ruihang Li¹Wei Li¹Xiaoding Ma^1*Yan Shen^1*Maomao Li²Binhua Hu³Ming Tang⁴Guomin Zhou¹Jian Wang¹Jianhua Zhang^1*

• ¹Chinese Academy of Agricultural Sciences (CAAS), Beijing, China
• ²Jiangxi Academy of Agricultural Sciences, Nanchang, China
• ³Sichuan Academy of Agricultural Sciences, Chengdu, China
• ⁴Heilongjiang Academy of Agricultural Sciences, Harbin, China

Key panicle traits, such as grain number, panicle length, and grain dimensions, are direct determinants of rice yield and market quality. As a staple food for over half of the world's population, rice plays a vital role in global food security, making the high-precision and high-throughput measurement of these traits critical for accelerating genetic improvement and modern rice breeding. In this study, 5300 high-quality rice panicle images (covering loose/normal/dense panicle types and milk/dough/full maturity/over-ripe stages) were used, with 3290 for training, 940 for validation, and 470 for testing. A deep learning model integrating object detection (OPG-YOLOv8), dynamic depth-first search (DFS) pruning algorithm, and linear regression was proposed, comprising modules for grain counting, panicle length extraction, grain length/width extraction, and maturity assessment. Experimental results showed that the panicle length extraction algorithm achieved a coefficient of determination (R²) of 0.9583, root mean square error (RMSE) of 5.69 mm, mean absolute error (MAE) of 4.91 mm, and mean absolute percentage error (MAPE) of 2.03%. For grain counting, a two-stage method combining density classification and regression correction effectively improved accuracy: R² values for loose, normal, and dense panicles were 0.9799, 0.9551, and 0.9278, respectively. Grain length extraction yielded an R² of 0.8823, MAE of 0.94 mm, and RMSE of 1.05 mm, while grain width extraction showed a MAPE of 6.64% with acceptable measurement precision. The OPG-YOLOv8 model outperformed comparative models (YOLOv8, YOLOv7, etc.) with a mAP50 of 99.10%, 59.6M parameters, and an inference speed of 154 FPS, balancing accuracy, computational complexity, and efficiency. Additionally, the yellowness value of rice panicle images (quantified via RGB normalization, gamma correction, and Hunter Lab parameter calculation) provided a reliable reference for maturity assessment, with an overall classification accuracy of 92.0%. A web-based "Rice Panicle Trait Extraction System" (developed via Django framework) was constructed, enabling one-stop service from image input to data export with an average processing time of 5 seconds per image. This study provides important technical support for rice breeding and promotes the intelligent and digital transformation of modern agricultural breeding.modern agricultural breeding.