Illegal timber trade is the most profitable natural resource crime, costing 50-152 billion USD a year with up to 90% of tropical timber in the supply chain being illegally sourced. Complex supply chains obscure timber sources, which often leaves traders’ declaration of origin as the sole evidence of provenance, despite the possibility of fraud. In recent years, there has been a growing interest in using chemical markers, such as Stable Isotope Ratios (SIRA) or Trace Elements (TEs) to verify claims of timber harvest location by matching levels of naturally occurring markers within wood tissue to location-specific values predicted from reference data (‘isoscapes’). However, overly simple statistical models for building isoscapes have so far limited the accuracy of and confidence in derived estimates of geographical origin. In this project, we will develop machine learning methods for inferring timber harvest location based on SIRA and TE data. We will use Gaussian Processes to robustly estimate isoscapes from reference wood samples, which will then be combined with species range data to compute, for every pixel in the study area, the probability of it being the origin of the sample. These models will be developed using timber data sets from World Forest ID - an international consortium that has sampled over 9000 trees from over 300 species globally. In a complementary line of research, we will design active learning (AL) methods to optimize future reference sampling campaigns by finding locations that maximize the expected accuracy of the derived isoscapes. The proposed methods will greatly advance the toolset available for identification of timber harvest origin, and will empower authorities worldwide in enforcing anti-deforestation legislation. Furthermore, our active learning approach will help maximize the utility of ongoing reference sampling campaigns carried out by World Forest ID, as well as any other sampling efforts.