Bridging soil mapping units and digital soil mapping through probability-driven machine learning: an uncertainty-aware framework for soil class prediction in Antarctica

Conventional mapping of soil classes has been performed in Antarctica since the early soil surveys. However, the subjectivity inherent to conventional mapping may yield many drawbacks, such as high uncertainties, lack of reproducibility and absence of means to assess map quality. These limitations have been successfully overcome by the integration of machine learning algorithms and digital soil mapping principles. In this work, we applied an approach with probability-based Random Forest to produce the first digital maps of soil classes for important ice-free areas of Maritime Antarctica and Northern Antarctic Peninsula. For this, we combined legacy soil data and a set of environmental covariates including soil physicochemical attributes, topographic and multispectral variables. The Random Forest achieved a mean accuracy of 0.43, kappa of 0.31, balanced accuracy of 0.61, and F1-score of 0.44, which can be considered a moderate to good level of performance when testing machine learning models with unseen data. The use of soil attributes as environmental covariates proved to be an approach with great potential for mapping soil classes, since their variability is directly related to soil formation processes in Antarctica. Based on the derivation of ranked soil class probabilities and their integration into homogeneous soil mapping units, we propose a novel digital soil mapping framework that reconciles machine learning–based probabilistic predictions with classical pedological concepts. Rather than representing soils as discrete and unequivocal entities following a deterministic modeling pathway, this approach explicitly incorporates spatial uncertainty and embraces the inherently probabilistic nature of digital soil mapping, by acknowledging the coexistence of multiple soil classes within a single mapped area. As a result, it provides a more realistic representation of soil spatial complexity in data-scarce Antarctic environments, while offering a conceptually robust and transferable basis for digital soil mapping in other remote and environmentally complex regions.