Abstract: This study addresses the challenge of maize and weed identification under varying lighting conditions, which often results in low detection accuracy and missed detections, particularly during the seedling stage of maize growth. To tackle this issue, a new detection method based on the WEED-YOLOv10 framework is proposed. The system was designed to enhance detection performance while maintaining computational efficiency. High-resolution field images were captured using UAVs to build a comprehensive dataset of maize and its associated weeds. The YOLOv10 architecture served as the baseline, but its backbone network was replaced with ConvNeXtV2, which is more capable of extracting detailed features from the input images. To further enhance the system’s robustness against lighting disturbances, the convolutional block attention module (CBAM) was integrated into the network. This module focuses attention on the most relevant features in the image, mitigating the impact of irrelevant information and improving model performance under diverse environmental conditions. Additionally, a SlimNeck structure was introduced to optimize the network’s computational efficiency, reducing unnecessary processing overhead while maintaining high feature representation capabilities. The incorporation of the Focaler-EIoU loss function further improved localization accuracy, ensuring precise identification of both maize and weed instances, even in challenging scenarios. Experimental results demonstrated that WEED-YOLOv10 outperformed the baseline model across several key evaluation metrics. The accuracy reached 85.4%, the recall rate was 88.1%, and the mean average precision (mAP) at an intersection over union (IoU) threshold of 50% (mAP@50) was 90.9%. The model also showed significant improvements in mAP at IoU thresholds ranging from 50% to 95% (mAP@50:95), with a score of 48.5%. The F1-score was 86.7%, reflecting the system’s ability to balance precision and recall. Compared to the baseline, the WEED-YOLOv10 model improved performance by 2.4%, 2.9%, 3.5%, 7%, and 2.6% across accuracy, recall, mAP@50, mAP@50:95, and F1-score, respectively. The model's inference speed was also highly optimized, achieving 28.7 frames per second when deployed on an NVIDIA Jetson Orin NX, ensuring real-time weed detection with a balance between speed and accuracy. In addition to the detection improvements, the system was integrated with a targeted pesticide spraying mechanism, allowing real-time capture and analysis of recognition signals. This integration enabled precise control of herbicide application based on the model’s output, ensuring that only weeds, not maize plants, were treated. Field tests demonstrated that the spraying system achieved a high spraying accuracy of 93.7%, a coverage rate of 90.5%, and a target deviation of just 0.0145m. The system was able to detect weeds at a speed of 20.1 frames per second, further demonstrating its suitability for automated weed control in maize fields. This study provides a significant contribution to the field of precision agriculture by offering a reliable, efficient solution for weed detection and management in complex lighting conditions. The proposed system enhances the speed, accuracy, and precision of weed control operations, making it a powerful tool for intelligent farming practices. The results suggest that the WEED-YOLOv10-based system can be a key enabler for the automation of field operations, paving the way for more sustainable, precise, and efficient agricultural practices, ultimately improving both productivity and resource management.