Label-Efficient Seedling Perception via Semi-Supervised Learning and Counterfactual Spatiotemporal Feature Rectification

In precision agriculture, assessing missing seedlings is essential for variable-rate reseeding and yield estimation. However, missing plants represent a semantic absence rather than explicit visual targets, making them difficult to define using conventional bounding-box annotations. To address this challenge, we propose a video-based, detection-driven gap inference framework for UAV wheat seedling phenotyping. Specifically, a semi-supervised Dual-Stream Anchor-Refinement Tracking (DART) mechanism is introduced to generate reliable pseudo-labels for visible seedlings from sparse temporal anchors, reducing dependence on dense manual annotations. To address visual fragmentation and weak target responses, a Masked Trajectory Inpainting Network (MTI-Net) combines object-aware masking with history-guided feature reconstruction to recover incomplete seedling representations using temporal context. The refined visible-seedling detections are then converted into agronomic gap indicators through a Dual-Scale Inertial Metric (DS-IM), which incorporates mechanical-sowing kinematic regularities and local growth-space constraints. Evaluations on the UAV-WheatSeedling dataset demonstrate positive mean gains when the proposed modules are embedded into RT-DETR and multiple YOLO-family backbones. When integrated with YOLO12-S, the method achieves 92.92±0.17% mAP@.50, 69.33±0.22% mAP@[.5:.95], and 86.59±0.41% Recall. Under 50% sparse-anchor supervision, the proposed framework approaches the 100% sparse-anchor YOLO12 baseline in mAP@[.5:.95], while achieving higher mean Recall. For downstream gap inference, DS-IM achieves an F1-score of 84.53% and a frame-level count MAE of 0.229 gaps/frame. This study provides a data-efficient and interpretable framework for low-altitude UAV-based early seedling assessment.