RGQ-DETR: Geometry-Conditioned Detection of Live and Dead Pepper, Tomato, and Cucumber Seedlings After Black Cutworm Exposure

Black cutworm (Agrotis ipsilon) is an underground pest whose larvae feed at the stem base of young crops, causing seedling wilt, severance, and stand loss. Assessment is difficult because larvae remain concealed in soil and injured plants can resemble healthy seedlings; dense planting, occlusion, and heterogeneous soil backgrounds further challenge detection. We construct an annotated live/dead seedling detection dataset for pepper, tomato, and cucumber seedlings and propose Representation–Geometry–Quality DETR (RGQ-DETR). Its P2-Guided Lodging Feature Recalibration (P2-LFR) conditions deeper channels on spatially weighted global context from high-resolution P2; Content-Gated Lodging Geometry Sampling (CG-LGS) adapts deformable sampling to candidate geometry; and Lodging Geometry-Quality Alignment (LGQA) refines candidate quality during training. RGQ-DETR achieves 59.76% AP and 72.83% AR100, improving over RT-DETR by 1.71 and 2.56 percentage points, respectively. For dead seedlings, it reaches 50.58% AP and 69.34% AR100, exceeding RT-DETR by 2.16 and 3.88 points. Compared with YOLO26m, DINO, and RF-DETR, it improves overall AP by 3.66, 3.86, and 4.25 points, respectively, with 77.89 M parameters. These results demonstrate its effectiveness for live/dead seedling monitoring under heterogeneous horticultural conditions.