An improved deep learning approach for detection of maize tassels using UAV-based RGB images
Jiahao Chen, Yongshuo H. Fu, Yahui Guo, Yue Xu, X. Zhang, Fanghua Hao
Abstract
The emergence of maize tassels is the turning of vegetative stage to reproductive stage of maize (Zea mays L.), which is critical for estimating maize grain yields. Recent advances in unmanned aerial vehicles (UAVs) remote sensing and deep learning-based object detection technique have provided a new approach for detecting maize tassels. Meanwhile, there still exists challenges for accurate detection due to the uncertainties in the complex field environment. The existing object detection networks fall in accurately detecting overlapping or small-scale maize tassels, as well as exhibiting insufficient detection capability in strong lighting conditions. Furthermore, the current dataset exhibits a limited temporal scope, unable to encompass the whole tasseling progress. In this study, we proposed FMTS dataset, designed a novel approach called RESAM-YOLOv8n (Residual Spatial Attention Module-You Only Look Once v8n), introducing the RESAM module and training the network with larger input image sizes. These enabled RESAM-YOLOv8n to focus on important tassel features and neglect irrelevant information, thereby enhancing its detection capability. The RESAM-YOLOv8n network was trained and evaluated using FMTS dataset, the mAP0.5, mAP0.75, Recall, Precision, and F1 of the network were 95.74 %, 66.70 %, 89.28 %, 95.59 %, and 92.00 %, respectively. Furthermore, in counting the number of maize tassels, the R2 value between the network’s detection and the ground truth reached 0.976, with a low RMSE of 1.56 tassels. The results showed the better performance of the RESAM-YOLOv8n network, providing an effective method for accurately identifying the maize tassels.