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

Marc Katzenmaier ^1,2* Vivien Sainte Fare Garnot ¹ Jan Dirk Wegner ¹ Georg von Arx ^2,3

• ¹ Department of Mathematical Modeling and Machine Learning, University of Zurich, Zürich, Switzerland
• ² Forest Dynamics, Swiss Federal Institute for Forest, Snow and Landscape Research (WSL), Birmensdorf, Zurich, Switzerland
• ³ Oeschger Center for Climate Research, Faculty of Science and Natural Sciences, University of Bern, Bern, Bern, Switzerland

Quantitative wood anatomy (QWA) along time series of tree rings (known as tree-ring anatomy or dendroanatomy) has proven to be very valuable for reconstructing climate and for investigating the responses of trees and shrubs to environmental influences. A major obstacle to a wider use of QWA is the time-consuming data production, which also requires specialized equipment and expertise. This is why the research community has been striving to reduce these limitations by defining and improving tools and protocols along the entire data production chain. One of the remaining bottlenecks is the analysis of anatomical images, which broadly consists of cell and ring segmentation, followed by manual editing, measurements and output. While dedicated software such as ROXAS can perform these tasks, its accuracy and efficiency are limited by its reliance on classical image analysis techniques. However, the reliability and accuracy of automatic cell and ring detection are key to efficient QWA data production. In this paper, we target automatic ring segmentation and deliberately focus on the most challenging case, circular ring structures in arctic angiosperm shrubs with partly very narrow and wedging rings. This shape requires high precision combined with a large global context, which is a challenging combination for instance segmentation approaches. We present a new iterative regression-based method for more precise and reliable segmentation of tree rings. We show a performance increase in mean average recall of up to 18.7 percentage points compared to previously published results on the publicly available MiSCS (Microscopic Shrub Cross Sections) dataset. The newly added uncertainty estimation of our method allows for faster and more targeted validation of our results saving a large amount of human labor. Furthermore, we show that panoptic quality performance on unseen species is more than doubled using multi-species training compared to single-species training. This will be another key step towards an AI-based version of the currently available ROXAS implementation.