Quantifying soil properties relevant to soil organic carbon biogeochemical cycles by infrared spectroscopy: The importance of compositional data analysis

Oxyhydroxides, soil texture and soil organic carbon (SOC) fractions are key parameters determining organic carbon cycling in soils. Standard laboratory methods to determine these soil properties are, however, time–consuming and expensive. Visible near infrared (Vis–NIR) and mid infrared (MIR) spectroscopy have been recognized as a promising alternative, but previous studies have not explicitly considered the above–mentioned soil properties as compositional data. The fractional components in compositional data are interrelated but their sum should be unity – these features should be represented in the spectral modeling process to minimize the prediction bias. In this study, two unique datasets were used to test these premises. The first one consisted of 655 samples collected from agricultural terraces and lynchets across Europe, which were scanned to acquire MIR spectra, while in the second one 4516 samples from private gardens across Flanders, Belgium were used to acquire Vis–NIR spectra. Memory–based learning models were optimized using both raw data (conventional method) and transformed data of soil properties by additive log–ratio (alr), centered log–ratio (clr), and isometric log–ratio (ilr) transformation methods. Results showed that the log–ratio transformation methods produced predictions as accurate as the conventional method, whilst also added two significant benefits: (1) they ensured the predicted fractions added up to 100% and (2) they reduced the number of samples with extreme prediction errors. We found that for 11 out of 18 investigated soil properties, the three log–ratio transformation methods provided similar model performance, whilst ilr outperformed clr for the prediction of silt and sand content of garden soils, for coarse particulate SOC (>250 µm) and microaggregate–associated SOC (250–53 µm) of terrace soils. For the remaining three properties (Al oxyhydroxides) alr outperformed ilr. Fair to excellent predictive models (RPD from 1.4 to 4.3; R2 from 0.50 to 0.95) were achieved for soil oxyhydroxides (Fe, Al, Mn) and soil texture from MIR spectra. Our approach also enabled accurate predictions of silt and sand content of garden soils using Vis–NIR spectra (RPD = 1.9; R2 = 0.72), although accuracy for clay was lower (RPD = 1.3; R2 = 0.49). This study demonstrates that combining soil infrared spectroscopy with a compositional data analysis is a promising technique that enables cost-effective and reliable quantification of soil properties relevant to SOC stability, thus offering a practical opportunity to assess the role of SOC in global C cycling.