Laboratory-based Vis–NIR spectroscopy and partial least square regression with spatially correlated errors for predicting spatial variation of soil organic matter content

Soil organic matter (SOM) has beneficial effects on soil properties for plant growth and production. Moreover,SOMchanges carbon dioxide concentrations in the atmosphere and can influence climate warming. Conventionalmethods for SOM determination based on laboratory analyses are costly and time consuming. Use ofsoil reflectance spectra is an alternative approach for SOM estimation and has the advantage of being rapid,non-destructive and cost effective. This method assumes that residuals are independent and identicallydistributed. However, in most cases this assumption does not hold owing to spatial dependence in soil samples.The aim of the paper was to test the potential of laboratory Vis–NIR spectroscopy to develop an approach ofpartial least square regression (PLSR) with correlated errors for estimating spatially varying SOM content fromlaboratory-based soil Vis–NIR spectra and producing a continuous map using a geostatistical method.The study area was the Turbolowatershed (Calabria, southern Italy),which is representative of Mediterranean areasbeing highly susceptible to soil degradation. Topsoil samples were collected at 201 locations. To reduce thelack of linearity that may exist in the spectra, reflectance (R) spectra were transformed in absorbance spectra(log (1 / R)). Partial least squared regression (PLSR) analysis was then used to predict SOM from reflectancespectra. To take into account spatial correlation between observations, the significant latent variables from PLSRwere used as regressors in a linearmixed effectmodel with correlated errors of SOM. The spatial approach and traditionalPLSRwere compared through the calculation of root mean square prediction error (RMSPE). In order to producea continuous map, the estimated SOMdata were interpolated by ordinary kriging. The approach is particularlyadvantageous when the data exhibit a pronounced spatial autocorrelation and could be used in digital soilmapping.