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

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