Reducing the Uncertainty in Determining the Origin of Natural Commodities via Earth Embeddings

Abstract

Verifying where commodities like timber were harvested is a forensic task, needed to detect illegal trade due to illegal deforestation or trade sanctions. Existing Gaussian-process models of wood chemistry give a Bayesian posterior over candidate harvest locations, but their credible regions in general do not deliver the coverage they promise (0.82-0.91 at the nominal 95\%). In this work we bring in a second, independent source of evidence: a prior over harvest locations learned from satellite (AlphaEarth) embeddings. We fit one classifier per genus that predicts presence from an embedding vector, supervised by occurrence records, and normalize its scores into a prior over the candidate grid. For such a prior to act on the prediction sets rather than merely mask cells, we build a cross-conformal predictor on the \emph{posterior cell mass}, so the prior moves every p-value while the coverage guarantee survives any misspecification of it. On four European genera, the embedding prior shrinks regions by up to 42\% compared to a uniform prior at no cost in coverage, which holds at 0.953-0.968 against a nominal 0.95.