Problem of dendrochronological measurement automation: modern approaches to solution

This article discusses the problem of automating dendrochronological measurements. Tree-ring analysis requires the preliminary determination of tree-ring boundaries, which is traditionally performed manually using a stereoscope, moving table, and data logger. However, this process is time-consuming, especially in the case of long tree-ring series. Developing a fully automatic, universal method for determining the boundaries of tree rings is a complex task due to variety of different tree species and types of their anatomical structures. There are several approaches to automatically detect ring boundaries; however, they use basic image processing methods (based on the first derivative of the image): Sobel filter, Canny edge detector, etc. As a result, their accuracy is limited, and their application is only possible for softwood where the boundaries of growth rings are clearly defined. There is also commercial software available, but none of them are universal because they do not work on ring-porous wood cores. Recent developments in machine learning increasingly prove that manual feature selection can be successfully replaced by automatic feature selection. Severalforeign works have begun to successfully use convolutional neural networks (CNN) in identifying, recognizing, and learning features, in which they have demonstrated fairly high performance. Moreover, since CNNs are able to identify the features which are beyond human perception, the learned features can surpass human accuracy in some applications. Moreover, their use does not require a priori knowledge about appearance of boundaries in tree rings, which makes the method universal for any type of wood. Their only shortage is the fact that training neural networks takes a lot of time. However, it must be remembered that this process needs to be done only once.