Maize is a monoecious cereal crop. Rapid and precise detection of maize tassels in the field is essential for irrigation management, pollination monitoring, and variety breeding. However, traditional one-stage object detection networks often suffer from performance decline under low-light conditions. To address this limitation, this study proposes YOLOv5-TT, a temperature-prior-guided one-stage deep learning framework based on YOLOv5 for accurate and efficient maize tassel detection. The proposed network incorporates a novel temperature feature-guided prior attention module (CATFC3M), which replaces the spatial attention component of the traditional CBAC3-1 module and significantly enhances feature representation through temperature feature-weighted mapping in low-light environments. Furthermore, three fusion methods are proposed based on the different fusion stages of RGB and thermal images: early fusion, mid-stage fusion, and late fusion. to determine the most effective integration stage for RGB and thermal information. Experimental results demonstrate that the proposed mid-stage fusion model achieves superior performance in natural field environments, with an average precision of 97.80%, an inference speed of 65.78 FPS, and a model size of 22.3 MB. Compared with conventional fusion approaches, the proposed temperature-prior-guided mechanism significantly improves detection robustness under low-light conditions. Overall, the proposed framework enables real-time maize tassel detection using fused RGB-thermal imagery and provides a reliable technical foundation for UAV-based agricultural monitoring.
Gao et al. (Sun,) studied this question.