Machine learning applied for Antarctic soil mapping: Spatial prediction of soil texture for Maritime Antarctica and Northern Antarctic Peninsula

Abstract

Soil texture is one of the most important soil properties as it drives several physical, chemical, biological, hydrological, and mechanical properties, and processes. In Antarctica the soil texture controls different important ecological processes, such as carbon stocks, nutrient leaching, and toxic metals retention. Thus, there is a pressing need for accurate soil data for the Antarctic ice-free areas, especially under the pressures imposed by climate changes and human disturbances on the continent. In this work, we predicted the distribution of sand, silt, and clay of the main ice-free areas of Maritime Antarctica and Northern Antarctic Peninsula, allowing novel soil texture maps to be produced for Antarctica. We used legacy soil texture data and tested different machine learning models, of which Random Forest presented the best performance to prediction. The best predictors were associated with climate, topography, and soil development degree, representing the scorpan model factors. The highest accuracy was obtained for clay prediction, mainly in the topsoil (LCCC of 0.49 and R2 of 0.31). In the mapping, the sand contents presented higher values, with a predominance of sandy loam texture, although clay contents are higher in some spots where chemical weathering is stronger. The final maps presented good spatial consistency, with soil texture distributed according to factors such as geomorphology, parent material and pedogenetic development. With our results, we aim to subsidize decision-making about soils in Antarctica, besides providing pioneer data that can be incorporated in global environmental models.