Assessing automatic ring detection on microscopy images of Salixglauca

Abstract

Shrub-ring analysis is increasingly used to assess climate-growth relationships in Arctic and alpine ecosystems. However, manual ring measurement remains labor-intensive and time-consuming, limiting the scale of ecological inference. To address this challenge, here we evaluated the Iterative Next Boundary Detection (INBD) deep learning method for automated ring detection using a new dataset of 50 manually annotated Salix glauca cross-section images from western Greenland. The model achieved intermediate performance, successfully detecting rings in morphologically clear samples but showing limited accuracy in more complex cases. We further evaluated three image resizing strategies and found that normalizing images to a fixed largest dimension of 1504 pixels improved segmentation accuracy and reduced training time compared to fixed downsampling approaches. We compared ring traces from automated and manual delineations, calculating basal area increment (BAI) from both approaches, along with six additional metrics derived from the manual ring traces. Growth patterns and ring counts from automated delineations were generally consistent with manually traced rings. Correlation analyses showed positive relationships between summer temperature and growth, with BAI (both automatic and manual) showing non-significant trends. In contrast, most one-dimensional metrics exhibited significant positive correlations, highlighting the potential influence of measurement approach on inferred climate sensitivity. Linear mixed-effects models further revealed consistent, significant positive relationships between shrub growth and mean summer temperature across all metrics, with the model based on automatically derived BAI explaining the largest proportion of variance. Our findings highlight both the potential and current limitations of the INBD method for automated shrub ring analysis in Salix glauca. Despite existing accuracy issues, the method can currently produce ecologically meaningful ring delineations and growth patterns. With species-specific training, refinement, and further testing, automated ring detection can accelerate data extraction from shrub rings and expand dendrochronological research in cold-climate regions.