Towards ROXAS AI: automatic multi-species ring boundaries segmentation as regression in anatomical images

Area of Science:

• Dendrochronology
• Quantitative Wood Anatomy (QWA)
• Image Analysis

Background:

• Quantitative wood anatomy (QWA) is crucial for climate reconstruction and understanding tree responses to environmental changes.
• Current QWA data production is time-consuming and relies on specialized equipment and expertise.
• Automated analysis of anatomical images, particularly cell and ring segmentation, remains a bottleneck.

Purpose of the Study:

• To develop a more precise and reliable method for automatic tree ring segmentation.
• To address the challenges of segmenting circular ring structures in arctic angiosperm shrubs, including narrow and wedging rings.
• To improve the efficiency and accuracy of QWA data production.

Main Methods:

• A novel iterative regression-based method was developed for automatic ring segmentation.
• The method was tested on the Microscopic Shrub Cross Sections (MiSCS) dataset, focusing on challenging circular ring structures.
• Uncertainty estimation was incorporated into the segmentation process.

Main Results:

• Achieved a performance increase of up to 18.7 percentage points in mean average recall compared to previous methods on the MiSCS dataset.
• The uncertainty estimation feature enables faster and more targeted validation, reducing human labor.
• Multi-species training more than doubled panoptic quality performance on unseen species.

Conclusions:

• The new iterative regression method significantly enhances the precision and reliability of tree ring segmentation.
• The developed method offers a substantial improvement in QWA data production efficiency.
• This work represents a key step towards an AI-based solution for QWA analysis, potentially replacing or augmenting existing software like ROXAS.