Segmentation for rice phenotype identification based on differential stratification LEGM

Accurate extraction of rice phenotypic features is the key to precision agriculture and efficient breeding. Existing models have made progress in rice coarse-grained recognition segmentation, but they cannot adapt to complex rice scenarios, lack robustness to backgrounds such as bare soil and field waterlogging, and are unable to distinguish normal rice from senescent vegetation or weeds. For this reason, the YOLO11-LEseg model is constructed in this paper. By embedding a multi-scale and hierarchical local feature enhancement module (LEGM) in the C3K2 module, a hierarchical modulation is formed to maintain the continuity of the feature flow and the stability of the gradient, and the design of “shallow view of the details, medium view of the fusion, and deep view of the whole world” is adopted to realize the multilevel feature characterization. Experiments on the RiceSEG dataset show that [email protected] and mIoU are improved by 3.1% and 4.5%, respectively, compared with the baseline YOLOv11m. The classification accuracy for five target types reached 100% at 0.956 confidence threshold. This method solves the core problems of rice phenotype segmentation, such as insufficient fine-grained feature capture, insufficient multi-scale fusion, and low tolerance of background interference, and provides technical support for high-throughput rice phenotype extraction.