Post by Jonas Schmidinger

Conventional laboratory analyses or soil spectroscopy? We argue, this is a false dilemma! Too often, soil testing is framed as: either we rely on conventional laboratory analyses, which are highly accurate but costly, or we use soil spectroscopy, which is cost-effective but often less reliable. Instead, we propose "reject-to-remeasure": a hybrid measurement framework for soil laboratories combining both approaches through probabilistic modelling with rejections. The idea is simple, a new soil sample is first analysed using soil spectroscopy. If the machine learning model predicts the soil property with sufficient confidence, the spectroscopic prediction is accepted. If the prediction is too uncertain, based on predefined quality thresholds, it is rejected and the sample is remeasured using conventional laboratory analysis. Using data from a soil laboratory in Québec (IRDA) in a cooperation led by Viacheslav Adamchuk, we demonstrate that this approach can combine the best of both worlds: reducing measurement costs through soil spectroscopy while maintaining accuracy through a conventional fallback option. Of course this approach requires reliable and conditional probabilistic modelling. In our study, we successfully used TabPFN and TabICL with scaled and tuned predictive distributions. This proved to be superior compared to common uncertainty approaches used in soil spectroscopy.