Spectral data mining for rapid measurement of organic matter in unsieved moist compost

Abstract

Fifty-five compost samples were collected and scanned as received by visible and near-IR (VisNIR, 350-2500 nm) diffuse reflectance spectroscopy. The raw reflectance and first-derivative spectra were used to predict log(10)-transformed organic matter (OM) using partial least squares (PLS) regression, penalized spline regression (PSR), and boosted regression trees (BRTs). Incorporating compost pH, moisture percentage, and electrical conductivity as auxiliary predictors along with reflectance, both PLS and PSR models showed comparable cross-validation r(2) and validation root-mean-square deviation (RMSD). The BRT-reflectance model exhibited best predictability (residual prediction deviation=1.61, cross-validation r(2)=0.65, and RMSD=0.09 log(10)%). These results proved that the VisNIR-BRT model, along with easy-to-measure auxiliary variables, has the potential to quantify compost OM with reasonable accuracy.