Abstract
Alfalfa (Medicago sativa), a globally important crop known for its high yields, wide adaptability, and high protein content, provides an excellent feed source for livestock. Smart breeding, an emerging technology that integrates genomics and phenomics, holds considerable promise for accelerating the development of elite varieties of alfalfa. Nevertheless, there are few phenotypic analysis tools available for alfalfa. Here, we present DU-Net-L, an effective and lightweight model for segmenting alfalfa images that enables preliminary analysis of branch phenotypes based on digital images. In our study, the DeepLabV3+ model struggled to handle petioles, while U-Net performed poorly with images captured under high light. To address these issues, we have created a new model utilizing ResNet34 as its feature extraction module and retaining the architectures of both DeepLabV3+ and U-Net. An analysis based on test data indicated that the new model overcame the shortcomings of using either of the two base models individually. Subsequently, we lightened the fused model by reducing output channels in each block, while maintaining its predictive capability. We have named the lightened model DU-Net-L. Ultimately, we adopted an exponential decay strategy for the learning rate and increased the number of training epochs to select an optimal parameter combination. This approach achieved 99.83% accuracy and a mean intersection over union of 0.9411, with a size of 25.42 MB. In summary, we have provided a lightweight model that effectively segments stems and leaves in alfalfa images, fulfilling the requirements for the preliminary analysis of branch phenotypes. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1007/s42994-025-00235-2.