Alfalfa-Yolo: a boundary-aware lightweight framework for real-time weed detection in alfalfa fields

Accurate weed detection in alfalfa fields is hindered by morphological similarity between crops and weeds at the seedling stage, substantial scale variation across weed targets, and degradation of fine spatial information during feature resampling. To address these challenges, Alfalfa-Yolo is proposed as a boundary-aware compact detection framework integrating three complementary components. The Edge-Perception Global Transfer Network (EPGTNet) uses bidirectional Sobel responses and cross-layer transmission to preserve edge-related high-frequency cues in deeper semantic features. The Multi-Scale Adaptive Feature Pyramid Network (MSAFPN) replaces static concatenation with attention-guided aggregation to adaptively regulate multi-scale contributions. The Enhanced Channel Shuffle Upsampling Module (ECSM) combines depthwise convolution, channel shuffle, and spatial shift to enhance local spatial representation after upsampling. On the self-constructed alfalfa-field dataset comprising 1,100 independent original field images, Alfalfa-Yolo achieves 77.6% mAP@0.5 and 55.9% mAP@0.5:0.95 with 2.67 M parameters. Independent-dataset benchmarking on CottonWeedDet12 and ImageWeeds confirms that the architecture remains competitive after target-dataset retraining, while a separate zero-shot cross-dataset protocol directly evaluates transfer without target-domain training. Deployment tests on NVIDIA RTX 4090, NVIDIA Orin NX, and RK3588 further characterize the accuracy-efficiency trade-off under different hardware and quantization settings. These results indicate that Alfalfa-Yolo provides a compact and practically deployable solution for real-time weed localization in alfalfa fields.