Abstract: Digital image processing techniques have been widely used to measure tree ring width, but most of which focus on conifer species with clear boundary. For hardwood species with complex anatomical structure and poor boundary of tree ring, traditional image processing techniques have poor performance. To improve the accuracy of tree-ring boundary recognition of broad-leaved tree species, we developed a tree-ring width measurement algorithm based on U-Net convolutional neural network model. An automatic tree-ring boundary recognition model based on U-Net convolutional neural network was constructed. Based on U-Net, we proposed a tree-ring boundary detection model for the tree cores of Picea jezoensis var. komarovii, Abies nephrolepis, Pinus koraiensi, Betula platyphylla, Betula costata, and Ulmus pumila. Three evaluation indices were used to compare the differences between U-Net method and manual labeling method, and the accuracy of tree ring width measured by WinDENDRO was compared. The results showed that tree ring boundary identified by U-Net accurately matches the actual boundary. Importantly, the detection accuracy of the tree ring boundary of broad-leaved trees was significantly improved compared with the traditional digital image processing method. The obtained tree ring width is proved to be accurate and reliable as evaluated by three evaluation indices, which has high practical value in tree ring analysis.

Key words: