Rice Seedling Leaf-Age Recognition Using Human-Inspired Active Separation for Imaging and YOLO-PFL

Rice seedling leaf age is important for determining transplanting time, but manual assessment is laborious, and leaf overlap and subtle differences between adjacent stages hinder automated recognition. This study aimed to improve leaf-age recognition in dense rice seedling nurseries by integrating active seedling separation with RGB and relative depth information. A device for human-inspired active seedling separation and close-up imaging was developed, and a dataset comprising 3077 images across six leaf-age categories was constructed. The proposed YOLO-PFL model was based on YOLO26n-seg and used four-channel early fusion of RGB images and spatially aligned relative depth maps estimated using Depth Anything V2. CKConvC3k2, C2PSA_CirculantAtt, and DPCF were incorporated to enhance morphological feature extraction for slender leaves, spatial relationship modeling, and multiscale detail fusion, respectively. Across four runs with random seeds 42, 43, 44, and 45, YOLO-PFL achieved mean test-set instance-mask precision, recall, F1 score, mAP@0.5, and mAP@0.5:0.95 of 91.42%, 77.61%, 83.95%, 86.05%, and 54.42%, respectively. Compared with YOLO26n-seg using the same four-channel input, these mean values increased by 3.57, 2.13, 2.75, 4.09, and 2.53 percentage points, respectively. These findings support the feasibility of class-aware instance segmentation for rice seedling leaf-age recognition under the evaluated greenhouse and post-separation imaging conditions.