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

Area of Science:

• Dendrochronology and Geomorphology
• Machine Learning Applications in Earth Sciences

Background:

• Dendrogeomorphology uses tree rings to date geomorphic events, but analyzing disturbed growth patterns is challenging.
• Deep learning (DL) excels at segmenting normal tree rings, yet its efficacy in disturbed environments is understudied.

Purpose of the Study:

• To evaluate the effectiveness of deep learning, specifically convolutional neural network (CNN)-based models, in segmenting tree rings with abnormal growth patterns relevant to dendrogeomorphology.
• To investigate the impact of image resolution, network architecture, and filtering techniques on CNN performance in disturbed tree-ring analysis.

Main Methods:

• Collected increment cores from a debris-flow-affected area.
• Acquired high-resolution images of tree rings and manually annotated boundaries and disturbances.
• Conducted experiments using various CNN architectures, image resolutions, and filtering techniques to assess segmentation accuracy.

Main Results:

• Segmentation performance decreased with growth disturbances showing significant color and texture variations.
• The framework successfully segmented narrow ring boundaries (>200 μm) when color was consistent, even in severely suppressed growth.
• Models using simple features like color variation performed similarly to those using finer cellular details; performance dropped without specified growth direction.

Conclusions:

• CNN-based methods show promise for dendrogeomorphological applications but have limitations with severe growth disturbances.
• Model performance is influenced by the degree of growth abnormality, image characteristics, and processing framework (e.g., growth direction specification).