Tree log identification using convolutional neural networks

Abstract

The identification of individual tree logs along the wood procurement chain is a coveted goal within the forest industry. The tracing of logs from the sawmill back to the forest would support the legal and sustainable sourcing of wood, as well as increase the resource efficiency and value of harvested timber. In this work, using a dataset of thousands of Scots pine (Pinus sylvestris L.) log end images displaying varying perspectives, lighting, and aging effects, we develop and assess log identification methods based on deep convolutional neural networks. The estimated rank-1 accuracy of our final model on an independent test set of 99 logs is 84 and 91% when allowing for random rotations of the log ends and when keeping each log at approximately fixed orientation, respectively. We estimate the scaling of these methods up to a template pool size of 493 logs, which reveals a weak dependence of accuracy on pool size for logs at fixed orientation. The deep learning approach gives superior results to a classical local binary pattern method, and appears feasible in practice, assuming that pre-filtering of the log database can be leveraged depending on the use case and properties of the queried log image. We make our dataset publicly available.