Spectral variable selection for estimation of soil organic carbon content using mid-infrared spectroscopy

The non-destructive and rapid estimation of soil total carbon (SOC) content with mid-infrared spectroscopy (MIR, 4000–400 cm−1) will play a vital role in precision agriculture. However, the benefit derived from the full MIR range is compromised by multicollinearity and noise. Hence, variable selection methods have been developed to reduce the full spectrum to a few variables that contribute the most information to a property of interest. However, only a few studies have applied variable selection methods in the MIR region to estimate organic carbon content in the soil. Therefore, four variable selection methods, namely stability competitive adaptive reweighted sampling (sCARS), bootstrapping soft shrinkage (BOSS), interval combination optimization (ICO) and the interval combination optimization-successive projections algorithm (ICO-SPA) method were investigated to ascertain the method that identified the wavebands most sensitive to SOC in order to improve prediction accuracy. The selected variables (i.e., reduced spectrum), as well as the full MIR spectrum, were coupled with partial least squares regression (PLSR) for the model calibration of SOC. The results showed that the models based on variable selection achieved higher prediction accuracy than the full spectrum model. sCARS selected 19 variables, BOSS selected 21 variables, ICO selected 311 variables, whereas ICO-SPA selected only 9 variables (accounting for 0.38% of all variables), while the prediction accuracy of the ICO-SPA-PLSR model was similar to those of the other three variable selection methods. The ICO-SPA-PLSR model had an Rp2 value of 0.93, RPD of 3.90 and RMSEP of 0.13%. This method identified 3450, 2920, 2767, 2000, 1800, 1765, 1600, 1560 and 927 cm−1 as the feature wavenumbers with the most useful information in accurately estimating SOC content. Therefore, the combination strategy of interval (ICO) and individual (SPA) variable selection may be a good alternative variables selection method for MIR spectroscopic data.