Field-Deployable Computer Vision Wood Identification of Peruvian Timbers

Abstract

Illegal logging is a major threat to forests in Peru, in the Amazon more broadly, and in the tropics globally. In Peru alone, more than two thirds of logging concessions showed unauthorized tree harvesting in natural protected areas and indigenous territories, and in 2016 more than half of exported lumber was of illegal origin. To help combat illegal logging and support legal timber trade in Peru we trained a convolutional neural network using transfer learning on images obtained from specimens in six xylaria using the open source, field-deployable XyloTron platform, for the classification of 228 Peruvian species into 24 anatomically informed and contextually relevant classes. The trained models achieved accuracies of 97% for five-fold cross validation, and 86.5 and 92.4% for top-1 and top-2 classification, respectively, on unique independent specimens from a xylarium that did not contribute training data. These results are the first multi-site, multi-user, multi-system-instantiation study for a national scale, computer vision wood identification system evaluated on independent scientific wood specimens. We demonstrate system readiness for evaluation in real-world field screening scenarios using this accurate, affordable, and scalable technology for monitoring, incentivizing, and monetizing legal and sustainable wood value chains.