Many wood items we import from abroad are incorrectly declared on customs forms. Sometimes it's an honest mistake, but other times, items made from illegally cut timber—commonly from tropical regions like the Congo and the Amazon—are deliberately mislabeled and enter the U.S. and other countries in violation of law. Accurate wood identification plays a critical role in maintaining sustainable wood product value chains and adhering to national laws such as the Lacey Act and international treaties like CITES.
Traditionally, wood identification has been the work of wood anatomists, specialists working in laboratories who examine wood with a magnifying glass or microscope looking for combinations of patterns in the wood's minute structure that are more or less unique to the species or genus. Today, FWRC scientist Dr. Frank Owens, an assistant professor in the Department of Sustainable Bioproducts, is finding promise in using artificial intelligence to identify wood. Dr. Rubin Shmulsky, sustainable bioproducts professor and head, and former doctoral student Adam Wade also are working on this project.
"The problem with the traditional method of wood identification is that there are not many wood anatomists, and training takes a long time," Owens said. "And customs agents inspecting wood product shipments typically don't have the skills to do it."
The XyloTron (the prefix "xylo" comes from the Greek x ýlon, meaning "wood") was invented by Dr. Alex Wiedenhoeft and a colleague at the USDA Forest Products Laboratory's Center for Wood Anatomy Research in Madison, Wisconsin. The system operates with a XyloScope, an instrument that connects to a laptop computer by USB cable and works like a microscope and high-resolution camera, and computer software using AI to analyze the picture and identify the species or genus. The operator prepares the end grain of a wood sample with a knife to reveal tiny anatomical features, takes a magnified picture of the cross-section, and pushes a button to start the image analysis and classification process, which takes only a few seconds. Wiedenhoeft and his research partner from the University of Wisconsin—Madison, Dr. Prabu Ravindran, design and customize the deep learning algorithms that allow the XyloTron to differentiate among different kinds of wood.
While they are not the only scientists investigating computer vision to identify wood species, the FWRC-based team is taking some innovative approaches to the practice using the XyloTron system.
"We focus on developing systems that can operate reliably in the field," Owens said. "One of the most important issues is image quality consistency. The XyloScope provides consistent lighting, framing, and distance from the lens to the wood, so you don't need to be in a controlled environment to capture quality images."
Wade, who focused his doctoral research on this project, added, "Similar systems require Wi-Fi or a cell phone signal, but with this software, you don't need connectivity on site."
While the speed and accuracy of the XyloTron are much faster and better than most humans, Owens and Wade acknowledge that it doesn't always identify specimens correctly, so there is still work to be done before the system can be fully commercialized as a timber identification tool for broad use.
"We are evaluating the ability of the computer model to make accurate predictions on wood samples it has never seen before," Wade said. He emphasized that testing the identification software only on wood samples used to train a model is not enough. "Accuracy commonly drops when you show it entirely new specimens."
In short, the tool is promising but not yet ready to replace human expertise. Having prepared and identified over 900 specimens for the study, Wade noted that the computer would also occasionally make surprisingly obvious mistakes in classification. "The more you ask it to do, the more mistakes it makes, and these are errors that a human wouldn't make," he said. "But we are still looking to challenge the XyloTron to do more difficult things and looking for alternative methods of training the AI software."
Breakthrough technology such as AI computer vision is complex and will not be perfected overnight. But it takes scientists like those from FWRC to set the foundation for further research that will radically transform the global fight against illegal logging in the future.
This research was funded by the Forest and Wildlife Research Center with partial funding from USDA ARS, the United States Department of State, the USDA Forest Service, the Forest Stewardship Council, and a Wisconsin Idea Baldwin Grant.