Tree ring segmentation using UNEt TRansformer neural network on stained microsections for quantitative wood anatomy

Area of Science:

• Forest ecology and dendrochronology
• Climate change impact on terrestrial ecosystems
• Plant anatomy and physiology

Background:

• Forests play a vital role in the global carbon cycle, making their response to climate change critical for future climate projections.
• Annual tree rings provide a historical record of environmental conditions and tree responses.
• Quantitative wood anatomy analyzes cellular structures for climate data, but manual tree ring delimitation is a bottleneck.

Purpose of the Study:

• To develop and evaluate an automated method for tree ring boundary delineation in stained wood microsections.
• To assess the accuracy and efficiency of neural networks compared to manual segmentation for quantitative wood anatomy.

Main Methods:

• Utilized a UNETR neural network, combining UNET and Visual Transformers, for segmenting annual ring boundaries.
• Trained the model on stained cross-sectional microsection images of beech wood cores.
• Evaluated segmentation accuracy against manual delineations and analyzed the impact on quantitative wood anatomy parameters.

Main Results:

• The automated UNETR model achieved high accuracy, matching or improving upon manual segmentation in 91.8% of cases.
• The assignment rate of vessels to annual rings was comparable between automated and manual methods.
• Neural network-based segmentation demonstrated superior performance over manual operators for specific quantitative wood anatomy analyses.

Conclusions:

• Automated tree ring boundary delineation using UNETR is a reliable and efficient alternative to manual methods.
• This approach can significantly reduce the cost and effort associated with large-scale, accurate data collection in quantitative wood anatomy.
• Improved data acquisition supports more robust analyses of forest responses to climate change.