Design of target spraying control system based on detecting maize and weed using WEED-YOLOv10

Maize and weed identification can often require for the high accuracy under varying lighting conditions, particularly at the seedling stage of the maize growth. In this study, an accurate and rapid detection was proposed to detect the corn and weed in field using WEED-YOLOv10 framework. Detection performance was then enhanced to maintain the computational efficiency. High-resolution images were captured from the field using UAVs. A dataset was then constructed for the maize and its weeds. The YOLOv10 architecture was served as the baseline. But its backbone network was replaced with the ConvNeXtV2, in order to extract the detailed features from the input images. Convolutional block attention module (CBAM) was integrated into the network, in order to further enhance the robustness against lighting disturbances. This module was also focused the attention on the most relevant features in the image. Irrelevant information was mitigated to improve the model performance under diverse environments. Additionally, a SlimNeck structure was introduced to optimize the computational efficiency of the network. Unnecessary processing was then reduced to maintain the high feature representation. Focaler-EIoU loss function was incorporated to improve the localization accuracy. Precise identification was realized on both maize and weed instances, even in challenging scenarios. Experimental results demonstrated that the WEED-YOLOv10 outperformed the baseline model over several key evaluation metrics. The high accuracy reached 85.4%, the recall rate of 88.1%, and the mean average precision (mAP) of 90.9% at an intersection over union (IoU) threshold of 50% (mAP@50). Significant improvements were achieved in the mAP at the IoU thresholds from 50% to 95% (mAP@50:95), with a score of 48.5%. The F1-Score was 86.7%, indicating the high performance to balance the precision and recall. Compared with the baseline, the WEED-YOLOv10 model was improved by 2.4%, 2.9%, 3.5%, 7%, and 2.6% over the accuracy, recall, mAP@50, mAP@50:95, and F1-Score, respectively. The inference speed was also highly optimized as 28.7 frames per second, when deployed on an NVIDIA Jetson Orin NX. The weed detection was obtained to balance the speed and accuracy in real time. In addition, the targeted pesticide spraying was integrated to capture the recognition signals. The herbicide application was precisely controlled using the output, in order to treat only weeds rather than the maize plants. Field tests demonstrated that the spraying system was achieved in a high spraying accuracy of 93.7%, a coverage rate of 90.5%, and a target deviation of only 1.45 cm. The weed was detected at a speed of 20.09 frames per second, suitable for the weed control in maize fields. A reliable and efficient solution can be offered for the weed detection under complex lighting conditions. The high speed, accuracy and precision of the weed control can greatly contribute to the field operations in intelligent farming. The WEED-YOLOv10 system can be expected for the more sustainable, precise and efficient agricultural practices. This finding can also provide the high productivity and resource management in precision agriculture.