Tree ring segmentation performance in highly disturbed trees using deep learning

Abstract

Dendrogeomorphology has provided valuable insights for dating geomorphic events, but requires the challenging analyses of tree-ring records from highly disturbed trees. Deep learning algorithms have been successfully used to detect ring boundaries under normal tree growth conditions. Here, we test if deep learning can perform tree ring segmentation in highly abnormal growth patterns. To this end, this study explores the relation between the complexity of convolutional neural networks (CNN)-based architectures, cellular detail levels, and the capacity to segment ring borders in abnormal tissues. Increment cores were collected from a debris flow-affected area in the Pyrenees, while images were acquired using a digital camera with a high-resolution macro. We defined four sets of experiments, including varying image resolution through downsampling, applying different architectures, and using image filters. Moreover, we test if the inclusion of the growth direction into a patchify-based system applied to increment cores improves the performance of the system. Our results suggest that intelligent systems can recognize tree-rings boundaries, but their performance was lower with high abnormal growth patterns due to the significant differences in colors and textures from normal growth patterns. However, the proposed segmentation system was able to segment sets of narrow ring borders, spaced above 200 m, where the color remained unchanged. Interestingly, our results suggest that the model ignored cellular details and relied on color gradients to detect ring borders when analyzing at the macro level. This implies that the image resolution is only becoming critical for densely packed rings with minimal spacing. Finally, we observed that CNN-based segmentation systems were unable to infer growth direction based solely on tree ring convexity and cellular details within an increment core patch. Our results provide new insights into how deep learning could be used in tree-ring research, but they still reveal the existing challenges with disturbed trees.