Reconstructing continental-scale variation in soil δ15N: a machine learning approach in South America

Abstract

Soil nitrogen isotope composition (δ15N) is an essential tool for investigating ecosystem nitrogen balances, plant–microbe interactions, ecological niches, animal migration, food origins, and forensics. The advancement of these applications is limited by a lack of robust geospatial models that are capable of capturing variation in soil δ15N (i.e., isotopic landscapes or isoscapes). Due to the complexity of the nitrogen cycle and general scarcity of isotopic information, previous approaches have reconstructed regional to global soil δ15N patterns via highly uncertain linear regression models. Here, we develop a new machine learning approach to ascertain a finer-scale understanding of geographic differences in soil δ15N, using the South American continent as a test case. We use a robust training set spanning 278 geographic locations across the continent, spanning all major biomes. We tested three different machine learning methods: cubist, random forest (RF), and stochastic gradient boosting (GBM). 10-fold cross-validation revealed that the RF method outperformed both the cubist and GBM approaches. Variable importance analysis of the RF framework pointed to biome type as the most crucial auxiliary variable, followed by soil organic carbon content, in determining the model performance. We thereby created a biogeographic boundary map, which predicted an expected multiscale spatial pattern of soil δ15N with a high degree of confidence, performing substantially better than all previous approaches for the continent of South America. Therefore, the RF machine learning framework showed to be a great opportunity to explore a broad array of ecological, biogeochemical, and forensic issues through the lens of soil δ15N.