Tree ring segmentation performance in highly disturbed trees using deep learning.

Dendrogeomorphology provides valuable insights into the dating of geomorphic events but requires complex analyses of tree-ring records from highly disturbed trees. While deep learning algorithms have been successfully applied to detect boundaries in normally developed growth rings, their performance under severely disturbed growth conditions remains largely unexplored. This study evaluates whether deep learning can effectively segment tree rings exhibiting abnormal growth patterns commonly observed in dendrogeomorphological contexts. Increment cores were collected from a debris-flow-affected area. High-resolution images were subsequently acquired and manually annotated to identify tree rings boundaries and growth disturbances. A series of experiments was conducted using different neural network architectures, image resolutions, and filtering techniques to examine the relationship between convolutional neural network (CNN)-based models and the level of cellular detail represented in the images. Our results indicate that segmentation performance declines in growth disturbances characterised by pronounced changes in colour and texture relative to normal growth patterns. Nevertheless, the proposed framework successfully identified sets of narrow ring boundaries spaced more than 200 μm apart when colour remained consistent, correctly segmenting most rings associated with the most severe growth suppressions in our dataset. Notably, models relying primarily on simple features such as colour variation performed comparably to those incorporating finer cellular details. We also found that, within a patch-based processing framework, performance decreased when growth direction was not specified in advance. Overall, this study provides a systematic evaluation of CNN-based methods under highly disturbed growth conditions, highlighting both their potential and current limitations in dendrogeomorphological applications.