CSAG-DETR: a lightweight detector for accurate weed detection in sugar beet fields under complex field conditions.

Introduction: Effective weed management is essential for reducing yield losses in sugar beet cultivation. However, existing deep learning-based detectors often experience performance degradation in complex field environments characterized by variable illumination, severe occlusion, dense vegetation, and background interference. In addition, models lacking lightweight designs generally require substantial computational resources, leading to increased inference latency and deployment costs that limit their application in real-time agricultural systems. Methods: To address these challenges, we developed CSAG-DETR, a lightweight and robust weed detection framework based on RT-DETR-R18 for accurate weed detection in sugar beet fields. The proposed framework incorporates a Cross-Stage Multi-Scale Network (CSMN) to enhance hierarchical feature interaction, a Cross-Stage Local-Detail Module (CSLM) to preserve fine-grained textures and object boundaries, and Global Attention-Gated Dual-Path Upsampling and Downsampling modules (GAGDU and GAGDD) to improve the stability of cross-scale feature transformation. Furthermore, we constructed the BeetWeed dataset containing nine common weed species and applied diverse data augmentation strategies to improve model robustness under variable field conditions. Results: Experimental results showed that CSAG-DETR achieved an mAP@0.5 of 98.9%, an mAP@0.5:0.95 of 78.6%, and an inference speed of 167.3 FPS on the BeetWeed dataset, outperforming 13 mainstream object detection models in terms of overall detection performance and computational efficiency. Generalization experiments conducted on the public CottonWeedDet12 dataset further demonstrated the strong cross-dataset generalization capability and competitive overall performance of the proposed model. Discussion: These results indicate that CSAG-DETR effectively balances detection accuracy, inference efficiency, and robustness in complex agricultural environments. The proposed framework therefore provides a practical solution for real-time weed detection and may support the deployment of intelligent weed management systems in sugar beet production.