Infrared spectroscopy prediction of soil microbial properties across Australian soils: Drivers and limitations

Soil microbial properties are increasingly recognised as a crucial metric of both soil health and soil carbon dynamics. However, these properties remain underutilised due to challenges in quantifying these properties at scale. This study assesses the potential for Near (NIR) and Mid (MIR) Infrared spectroscopy, as well as Granger Ramanathan Averaging models that integrate predictions from both spectra to provide reliable estimations of microbial properties based on 120 soil samples collected across Australia. Given that these properties are not spectrally active, the accuracy of their predictions relies on correlations with physicochemical properties such as organic matter, clay minerals and soil moisture. Using soil samples from across Australia, the most effective models for DNA yield, microbial respiration and N-acetyl-β-D-glucosaminidase enzyme activity achieved R2 of 0.7, 0.63 and 0.62, though microbial biomass carbon (MBC) and β-glucosidase enzyme activity predictions were more modest with R2 = 0.52 and 0.49, respectively. MIR consistently outperformed NIR, whereas data fusion models were comparable to MIR across all properties, except for a slight improvement in MBC from R2 = 0.46 to R2 = 0.52. From the correlational analysis, the key properties driving these predictions were moisture for both enzymes (r = 0.35-0.36) and organic carbon for all other properties (r = 0.34-0.56), as indicated by significant wavenumbers in the MIR spectrum. However, the reliance on the moisture peak highlights a potential limitation in predicting properties measured from field soils using spectra generated from processed, air-dried samples. Overall, these findings are generally encouraging for predicting microbial properties from the MIR spectrum, supporting their future use in assessing soil health and monitoring carbon dynamics. Verifying these correlations is crucial to ensuring the wider applicability of these algorithms across diverse samples and to expanding spectral libraries.