Method for detecting tree ring boundary in conifers and broadleaf trees using Mask R-CNN and linear interpolation

Abstract

As tree rings can reveal various information regarding climate and environmental factors, increasing research is being conducted on them. Although tree ring analysis software such as Windendro has been applied, research on the development of analysis software using image preprocessing algorithms and deep learning is recently being attempted as computer vision technology. In this study, Mask R-CNN and linear interpolation were applied from images collected using a smartphone (SM-G973, Samsung, Suwon) and a scanner (CanoScan 9000F, Canon, Otaku) to propose an effective method for detecting tree ring boundaries. Pitch pine (Pinus rigida), Korean pine (Pinus koraiensis), white birch (Betula platyphylla), and cork oak (Quercus variabilis) were selected as tree species. Of the 300 images, 240 were classified as training data and 60 as validation data. As a result of learning, smartphones detected 86.0 % (381 ring boundaries) of the rings in pitch pine, 82.1 % (367 ring boundaries) in Korean pine, 84.7 % (309 ring boundaries) in white birch, and 78.7 % (318 ring boundaries) in cork oak. The scanner detected 93.2 % (413 ring boundaries) of the rings in pitch pine, 90.8 % (405 ring boundaries) in Korean pine, 88.2 % (322 ring boundaries) in white birch, and 89.4 % (361 ring boundaries) in cork oak. In particular, the smartphone showed satisfactory results of 84.7 % and 78.7 % for detecting tree ring boundaries of white birch and cork oak, where the boundaries of the rings were unclear. In the annual growth analysis results, both smartphones and scanners were statistically insignificant, and there was no difference compared with those of Windendro. Therefore, Mask R-CNN might be an effective approach for tree ring boundary detection as it showed satisfactory results, even with smartphones. In addition, although there was distortion in cases where images were acquired with a circular lens, there was no statistically significant difference from Windendro results. Thus, Mask R-CNN and linear interpolation can be used for tree ring boundary detection and growth measurement.