The monitoring of the tassel number and tasseling time reflects the maize growth and is necessary for crop management. However, it mainly depends on field observations, which is very labor intensive and may be biased by human errors. Tassel detection remains challenging due to the varying appearance of tassels across maize varieties, tasseling stages, and spatial resolutions. Moreover, the capability of the deep learning model for monitoring tassel number change and the time of entering tasseling stage has not been explored. In this study, we propose a novel approach for fast tassel detection using PConv (Partial Convolution) within YoloV8 series, named PConv-YoloV8 series. Compared to seven state-of-the-art deep learning methods, PConv-YoloV8 × 6 best trades off detection accuracy with the number of parameters (Parameters = 52.50 MB, AP = 0.950, R² = 0.92, rRMSE = 9.08%). The potential of PConv-YoloV8 × 6 to provide an accurate detection of tassels in complex situations from near-ground and UAV images were comprehensively studied. PConv-YoloV8 × 6 maintained an excellent detection accuracy for maize at different tasseling stages (AP = 0.826–0.972, R² = 0.83–0.92, RMSE = 1.94–3.01, rRMSE = 21.06%-7.09%), for different varieties (AP = 0.901–0.978, R² = 0.77–0.97, RMSE = 1.39–3.16, rRMSE = 11.72%-5.06%), at different resolutions (AP = 0.921–0.956, R² = 0.84–0.93, rRMSE = 8.72%-17.71%), and on UAV images with different resolutions (AP = 0.918–0.968, R² = 0.98–0.99, rRMSE = 6.43%-12.76%), which proved the robustness of the model. The tasseling number and the time of entering tasseling stage detected from images were basically consistent with the trends observed in the manually labeled results. This study provides an effective method to monitor the tassel number and the time of entering the tasseling stage. A new maize tassel detection dataset (18260 tassels in 729 near-ground images and 20835 tassels in 144 UAV images) is created. Future studies will focus on making more lightweight models and achieving real-time detection capabilities.

雄穗数量和抽雄时间的监测反映了玉米的生长情况，对于作物管理是必要的。然而，它主要依赖于现场观察，这是非常劳动密集型的，并且可能会因人为错误而产生偏差。由于不同玉米品种的雄穗外观、抽雄阶段和空间分辨率各不相同，雄穗检测仍然具有挑战性。此外，深度学习模型监测雄穗数量变化和进入抽雄阶段时间的能力尚未被探索。在本研究中，我们提出了一种使用 YoloV8 系列中的 PConv（部分卷积）进行雄穗快速检测的新方法，称为 PConv-YoloV8 系列。与七种最先进的深度学习方法相比，PConv-YoloV8 × 6 最好地权衡了检测精度与参数数量（参数 = 52.50 MB，AP = 0.950，R ²  = 0.92，rRMSE = 9.08%）。全面研究了 PConv-YoloV8 × 6 在复杂情况下从近地和无人机图像中准确检测雄穗的潜力。 PConv-YoloV8×6对不同抽雄阶段的玉米保持了优异的检测精度（AP = 0.826–0.972，R ²  = 0.83–0.92，RMSE = 1.94–3.01，rRMSE = 21.06%-7.09%），对于不同品种（AP = 0.901–0.978，R ²  = 0.77–0.97，RMSE = 1.39–3.16，rRMSE = 11.72%-5.06%），不同分辨率（AP = 0.921–0.956，R ²  = 0.84–0.93，rRMSE = 8.72%-17.71 %），并在不同分辨率的无人机图像上（AP = 0.918–0.968，R ²  = 0.98–0.99，rRMSE = 6.43%-12.76%），证明了模型的鲁棒性。图像检测到的抽雄数量和进入抽雄阶段的时间与人工标记结果观察到的趋势基本一致。本研究为监测雄穗数量和进入抽雄阶段的时间提供了有效的方法。创建了新的玉米雄穗检测数据集（729 个近地图像中的 18260 个雄穗和 144 个无人机图像中的 20835 个雄穗）。未来的研究将集中于制作更轻量级的模型并实现实时检测功能。