【目的】针对当前水稻育种过程中人工测量始穗期存在劳动强度大、效率低等问题，本研究提出一种融合无人机图像、YOLOv10b目标检测算法与随机森林回归模型的智能预测方法，实现水稻育种小区始穗期的自动化高效预测。【方法】以水稻育种小区为研究对象，利用无人机在水稻孕穗至成熟阶段获取15个观测期的小区图像，基于YOLOv10b算法训练稻穗检测模型，实现小区图像中稻穗的识别与计数。随后以小区各观测期的稻穗数量为输入变量，始穗期距移栽天数为预测变量，构建多种始穗期预测模型并进行对比分析。在此基础上开展特征重要性分析，进行特征降维，构建精简模型。【结果】YOLOv10b模型在稻穗检测任务中精确率为91.7%，召回率为92.1%，F₁值为0.919。在始穗期预测任务中，基于全部15个观测期穗数时序特征所构建的随机森林模型取得了R²为0.84，RMSE为1.57 d，MAE为1.12 d的预测表现；进一步通过特征重要性降维至9个观测期所构建的精简模型，取得了R²为0.83，RMSE为1.65 d，MAE为1.16 d的预测表现，在保证预测精度的同时，显著降低了特征维度，提升了模型的效率与实用性。【结论】该研究提出的方法可有效减少人工测量的工作量，提升水稻始穗期监测效率，实现育种小区始穗期的自动化预测，为智能育种与精准农业管理提供可行的数据支撑与技术路径。

【Objective】To address the problems of high labor intensity and low efficiency in manual measurements of heading date in rice breeding, this study proposes an intelligent prediction method that integrates UAV imagery, the YOLOv10b object detection algorithm, and a random forest regression model to achieve automated and efficient prediction of heading date in rice breeding plots.【Methods】Using rice breeding plots as the research subject, UAV imagery was acquired at 15 observation time points from booting to maturity. Based on the YOLOv10b algorithm, a rice panicle detection model was trained to recognize and count panicles in plot images. Subsequently, the numbers of panicles at each observation time point were used as input variables, and the number of days from transplanting to heading was used as the prediction variable to construct and compare multiple heading date prediction models. Based on this, feature importance analysis and dimensionality reduction were performed to develop a simplified model.【Results】Experimental results showed that the YOLOv10b model achieved a precision of 91.7%, a recall of 92.1%, and an F₁ score of 0.919 in the panicle detection task. In the heading date prediction task, the random forest model constructed using the sequential features of panicle counts from all 15 observation time points achieved an R² of 0.84, an RMSE of 1.57 days, and an MAE of 1.12 days. Furthermore, after feature dimensionality reduction based on feature importance analysis, a simplified model constructed using 9 observation time points achieved an R² of 0.83, an RMSE of 1.65 days, and an MAE of 1.16 days. This approach effectively reduced feature dimensionality while maintaining prediction accuracy, thereby improving the model's efficiency and practicality.【Conclusion】The proposed method can effectively reduce the workload of manual measurements, improve the efficiency of heading date monitoring, and realize automated prediction of heading date in breeding plots. It provides a feasible data foundation and technical pathway for intelligent breeding and precision agricultural management.