Soil bacterial abundance and diversity better explained and predicted with spectro-transfer functions. (February 2019)

• Record Type: Journal Article
• Authors: Yang, Yuanyuan; Viscarra Rossel, Raphael A.; Li, Shuo; Bissett, Andrew; Lee, Juhwan; Shi, Zhou; Behrens, Thorsten; Court, Leon
• Is Part Of: Soil biology and biochemistry. Volume 129(2019)
• Page Start: 29
• Page End: 38
• Publication Date: 2019-02
• Subjects: Soil bacteria -- Bacteria diversity -- Bacteria phyla -- Machine learning -- Vis–NIR spectra -- Spectro-transfer function; Soil biochemistry -- Periodicals; Soil biology -- Periodicals; Sols -- Biochimie -- Périodiques; Sols -- Biologie -- Périodiques; Sols -- Microbiologie -- Périodiques; Bodembiologie; Biochemie; 631.46
• DOI: 10.1016/j.soilbio.2018.11.005
• Languages: English
• ISSNs: 0038-0717

Abstract

Abstract: Soil bacteria play a critical role in the functioning of ecosystems but are challenging to investigate. We developed state-factor models with machine learning to understand better and to predict the abundance of 10 dominant phyla and bacterial diversities in Australian soils, the latter expressed by the Chao and Shannon indices. In the models, we used proxies for the edaphic, climatic, biotic and topographic factors, which included soil properties, environmental variables, and the absorbance at visible–near infrared (vis–NIR) wavelengths. From a cross-validation with all observations (n = 681), we found that our models explained 43–73% of the variance in bacterial phyla abundance and diversity. The vis–NIR spectra, which represent the organic and mineral composition of soil, were prominent drivers of abundance and diversity in the models, as were changes in the soil-water balance, potential evapotranspiration, and soil nutrients. From independent validations, we found that spectro-transfer functions could predict well the phyla Acidobacteria and Actinobacteria ( R 2 > 0.7) as well as other dominant phyla and the Chao and Shannon diversities ( R 2 > 0.5). Predictions of the phyla Firmicutes were the poorest ( R 2 = 0.42). The vis–NIR spectra markedly improved the explanatory power and predictability of the models. Highlights: We developed machine learning models of soil bacteria phyla abundance and diversity. The models explained 43–73% of the variance in abundance and diversity. Drivers varied but vis–NIR, soil nutrients and climate were generally most important. Spectro-transfer functions could predict bacterial phyla abundance and diversity. Vis–NIR spectra improved the explanatory power and predictability of the modelling by up to 39%.