Detecting tasseling rate of breeding maize using UAV-based RGB images and STB-YOLO model

In the regions with limited light and temperature, detecting the tasseling rate of maize is crucial to optimize water and fertilizer management, adjusting harvest schedule and screen suitable varieties. Unmanned Aerial Vehicle (UAV) imaging technology offers a rapid method for detecting the maize tasseling rate. This study proposes a new detection model, STB-YOLO, based on YOLOv8 for detecting maize tasseling rate. At first, we introduced Swin Transformer blocks in the downsampling process to enhances the ability of semantic feature extraction from UAV-based RGB images. Subsequently, the Bidirectional Feature Pyramid Network is employed during the Concat fusion process. This enhances the model's ability to accurately detect and robustly handle targets of varying scales in images with different tasseling rate. Finally, a series of deep learning algorithms are compared and analyzed. Additionally, the model is analyzed in detail by ablation experiment. The results show that at imaging heights of 15 meter and 30 meter, STB-YOLO achieved a precision of 76.2% and 72.1%, respectively. This shows an improvement of 6.5 and 11.7 percentage over YOLOv8 and YOLOv6, respectively. The precision of tasseling rate in the test datasets reaches 78.48% and 73.22%, with R² of 0.71 and 0.69, respectively. The precision increases as the tasseling rate increases. When the tasseling rate exceeds 80%, the precision reaches 93.44% and 87.01%, respectively. Therefore, applying the STB-YOLO deep learning algorithm to UAV imagery facilitates accurate detection of tasseling rates of breeding maize.