Predicting soil physical and chemical properties using vis-NIR in Australian cotton areas

Abstract Management of Vertosols in southeast Australia, requires information about soil physical (e.g. particle size fractions) and chemical (e.g. cation exchange capacity [CEC – cmol(+) kg−1], exchangeable sodium percentage [ESP - %] and pH) properties. While visible and near-infrared (vis-NIR) spectroscopy calibration models have been developed, little has been done in Vertosols. The performance of multi-depth or depth-specific (i.e. topsoil [0–0.3 m], subsurface [0.3–0.6 m] and subsoil [0.9–1.2 m]) calibration models has also seldom been discussed. In this paper, using a spiking approach across seven cotton growing areas, our first aim was to determine which model (e.g. machine learning algorithm (Cubist) or partial least square regression with bootstrap aggregation [bagging-PLSR]) produced better calibrations using multi-depth data. The second aim was to see how these calibrations predict depth-specific soil properties using independent validation. Our third aim was to investigate whether depth-specific calibrations could produce better predictions. In terms of multi-depth calibration, exemplified by CEC, Cubist (R2 = 0.86) was stronger than bagging-PLSR (0.72). However, in terms of prediction agreement for independent validation, bagging-PLSR was superior to Cubist in the topsoil (LCCC = 0.84) and subsoil (0.83) and equivalent in the subsurface (0.74). Moreover, the depth-specific bagging-PLSR achieved equivelent prediction agreement for the independent validation of CEC to the multi-depth bagging-PLSR in the topsoil (LCCC = 0.85), subsoil (0.85) and subsurface (0.76). In terms of the other soil properties (i.e. clay, silt and sand), multi-depth bagging-PLSR was superior and overall a multi-depth spectral library is recommended for Vertosols. This has implications for acquiring a vis-NIR library more quickly and prediction efficiency with multi-depth calibrations.