Probabilistic boosting models for accurate timber tracing An alternative spatial approach for tracing timber samples by using stable isotope ratios and trace elements analysis

Abstract

Illegal logging is estimated to generate between 52 and 157 billion USD per year according to a 2017 Global Financial Integrity report, with severe environmental, social, and economic consequences. By falsifying timber certificates of origin, fraudulent actors are able to circumvent law enforcement and trade regulations. Therefore, tools capable of identifying the true origin of timber are crucial for preventing the trade of illegally harvested timber. In this thesis, we propose the use of natural gradient boosting (NGBoost), a probabilistic boosting method, as a replacement for the Gaussian process regressors used in the current state-of-the-art timber tracing spatial determination approach. We evaluate the proposed method against the established methodology using datasets containing stable isotope and trace element data from timber originating in the US and Europe. The proposed approach achieves comparable mode distance and lower absolute errors, while offering additional modeling flexibility, such as the possibility to use different base learners and to model the chemical markers with different probability distributions, including multivariate ones, making it a viable and adaptable alternative to current timber tracing methods.