Compositional Data Methods and VISNIRS to Predict Soil Organic Carbon Contents

Soil organic carbon (SOC) content plays an important role in modulating atmospheric CO 2 . Visible and near‐infrared spectroscopy (VISNIRS) has been proven to be a suitable method for SOC prediction in the laboratory. However, several soil properties such as soil moisture (SM), bulk density, compactness, texture, and temperature affect the near‐infrared spectra obtained under field conditions. Among these factors, SM variation is the most significant challenge for SOC measurement. Soil is a composition of fractions, especially minerals and organic matter, whose contents are expressed in relative and interdependent quantities, belonging to simplex spaces. These are known as compositional data (CoDa) and require specific mathematical methods. This study proposes methods to predict SOC along with other soil components, rather than using solely one soil feature. Several predictive models using VISNIRS by considering different soil compositions were evaluated. All models included SM to mitigate its interference in SOC prediction, which would otherwise occur when using only VISNIRS‐based methods. The analyzed soil components included soil organic matter (SOM, calculated as SOM = 1.724 × SOC), SM, soil inorganic carbon (SIC), and the textural fractions: “Clay,” “Silt,” and the remainder of the soil sample classified as “Other.” The 4‐parts model including the clay content provided SOM prediction with Lin's concordance correlation coefficient = 0.84 and Pearson r = 0.87. Important is to note that the predictions stated with the different CoDa approaches showed similar trends, from the 6‐Parts to the 2‐Parts compositions, this fact highlighting the consistency of the method. The performance of all the CoDa models obtained, and in particular the 4‐part “Clay” model, was superior to that obtained with the traditional PLS calibration. The results highlighted that CoDa methods for estimating SOM or SOC provided an improvement over traditional partial least square (PLS) calibration. Future software solutions could integrate routines for using these methods in the field.