Rapid field identification of cites timber species by deep learning

• Published in: Trees, Forests and People (Online) Vol. 2; p. 100016
• Main Authors: Olschofsky, Konstantin, Köhl, Michael
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.12.2020; Elsevier
• Subjects: CITES; Convolutional neuronal networks; Deep learning; Forest protection; Illegal logging; legal timber trade; Deep learning; Forest protection; Convolutional neuronal networks; legal timber trade; CITES; Illegal logging
• ISSN: 2666-7193; 2666-7193
• DOI: 10.1016/j.tfp.2020.100016

Enforcement of the ban on trade in protected tree species is often hampered by insufficient knowledge of species identification by the authorities responsible for control. We present a new method based on a deep learning approach that facilitates the separation of protected and non protected tree species. The classification system used here is based on the image classification model Inception-v3, a so-called transfer learning tool provided by Google. Inceptionv3 uses a convolutional neural network (CNN) for feature extraction and classification and has been pre-trained with 1.2 million images. Input and bottleneck layers of the pre-trained CNN are used to generate new CNNs to which images are passed and classified. As an example of a protected tree species, we have chosen Cedrella odorata, which we compare with 13 other tropical tree species. In a first step, the data set of C. odorata and five other unprotected tree species was partitioned into five groups using a bootstrap algorithm. Then the images of each group were passed to an image classification system and used to generate five CNNs independently. Each CNN assigns image data to the two classes of protected tree species, i.e. C. odorata, or unprotected tree species. The five CNNs were first verified with images of the same six tree species that were used for training. For this purpose, images were taken that were not used for CNN training. 98 percent of the image scans were correctly classified into the two classes of protected and non-protected species. In a second verification step, additional tree species not used for training were presented to the image classification system. Here the correct allocation was reduced to 87 percent. Both the tree species and the CNNs used influence the accuracy of the correct class assignment. Against the background that only a few samples were used to train the CNNs, the accuracy achieved is convincing and shows the potential of simple applicable pre trained deep learning for the operational field inspection of timber loads as a first step to enforce legal timber trade regulations.