Abstract: Dendrochronological research uses tree-age and annual-ring width to estimate environmental changes and tree growth. Thus, it is important to accurately extract the characteristics such as the early wood, late wood, and bark parts in the annual-ring images for further analysis. It is difficult to obtain the desired effect using traditional image segmentation algorithm due to the existence of defects such as fuzzy interface between the early and late woods, knots and pseudo-annual rings during growth and there are burrs and noise spots on the image of the annual ring disc during the cutting and collecting process. Here, we proposed a novel approach to perform annualring image semantic segmentation based on convolutional neural network. Firstly, 100 annual-ring images were marked as late wood, bark and other parts. Data enhancement was implemented- through image rotation, perspective, and deformation to generate 20000 image data, from which 16000 images were randomly selected as training data sets and 4000 images were used as test dataset. Secondly, according to the characteristics of image dataset, an annual-ring disc image segmentation network was developed based on U-Net convolutional network using the Tensorflow framework. Then, the training dataset was sent into the network, the training parameters were optimized, and the annual-ring image segmentation network was iteratively trained until the evaluation index and the loss function no longer change. Finally, the test dataset was segmented using the trained model and the segmentation indicators were evaluated. Experimental results showed that the constructed model can effectively avoid the defects mentioned above, and completely separate the late wood and bark parts of the annual-ring images. The proposed approach was tested with dataset consisting of 4000 tree ring images, the corresponding accuracy of mean pixels and the mean intersection over union achieved 96.51% and 82.30%, respectively. Thisapproach based on U-Net convolutional network is a more efficient algorithm for annual-ring image segmentation, with stronger generalization ability and robustness.

Key words: spatial heterogeneity, coastal saline area, salt-tolerant plant, soil chemical factor