Estimating soil fungal abundance and diversity at a macroecological scale with deep learning spectrotransfer functions

Abstract

Abstract. Soil fungi play important roles in the functioning of ecosystems, but they are challenging to measure. Using a continental-scale dataset, we developed and evaluated a new method to estimate the relative abundance of the dominant phyla and diversity of fungi in Australian soil. The method relies on the development of spectrotransfer functions with state-of-the-art machine learning and uses publicly available data on soil and environmental proxies for edaphic, climatic, biotic and topographic factors, and visible–near infrared (vis–NIR) wavelengths, to estimate the relative abundances of Ascomycota, Basidiomycota, Glomeromycota, Mortierellomycota and Mucoromycota and community diversity measured with the abundance-based coverage estimator (ACE) index. The algorithms tested were partial least squares regression (PLSR), random forest (RF), Cubist, support vector machines (SVM), Gaussian process regression (GPR), extreme gradient boosting (XGBoost) and one-dimensional convolutional neural networks (1D-CNNs). The spectrotransfer functions were validated with a 10-fold cross-validation (n=577). The 1D-CNNs outperformed the other algorithms and could explain between 45 % and 73 % of fungal relative abundance and diversity. The models were interpretable, and showed that soil nutrients, pH, bulk density, ecosystem water balance (a proxy for aridity) and net primary productivity were important predictors, as were specific vis–NIR wavelengths that correspond to organic functional groups, iron oxide and clay minerals. Estimates of the relative abundance for Mortierellomycota and Mucoromycota produced R2≥0.60, while estimates of the abundance of the Ascomycota and Basidiomycota produced R2 values of 0.5 and 0.58 respectively. The spectrotransfer functions for the Glomeromycota and diversity were the poorest with R2 values of 0.48 and 0.45 respectively. There is no doubt that the method provides estimates that are less accurate than more direct measurements with conventional molecular approaches. However, once the spectrotransfer functions are developed, they can be used with very little cost, and could serve to supplement the more expensive and laborious molecular approaches for a better understanding of soil fungal abundance and diversity under different agronomic and ecological settings.