Laser-Induced Breakdown Spectroscopy (LIBS) for tropical soil fertility analysis

The Laser Induced Breakdown Spectroscopy (LIBS) is a promising technique for soil fertility analysis in a rapid and environmentally friendly way. This application requires the selection of an optimal modelling procedure capable of handling the high spectral resolution of LIBS. This work aimed at comparing different modelling methods of LIBS data for the determination of key fertility attributes in Brazilian tropical soils. A benchtop LIBS system was used for the analysis of 102 soil samples, prepared in the form of pressed pellets. Models for the prediction of clay, organic matter, pH, cation exchange capacity, base saturation, and the extractable nutrients P, K, Ca, and Mg were developed using univariate linear regression (ULR), multiple linear regression (MLR) and partial least squares regression (PLS). The following input data for PLS were used: (i) the full spectra from 200 to 540 nm (38,880 variables), and (ii) variables selected by the interval successive projections algorithm (iSPA). The multivariate models achieved satisfactory predictions [residual prediction deviation (RPD) > 1.40] for eight out of nine fertility attributes. However, the best performances were obtained for the PLS with the variable ranges selected by the iSPA, which achieved satisfactory predictions (RPD ≥ 1.44) for seven out of the nine soil attributes studied. The MLR method obtained lower prediction performance than the iSPA-PLS using only 21 variables. The iSPA-PLS approach allowed a reduction from 3 to 160-fold in the total of variables compared to the full LIBS spectra, making it efficient and accurate modelling method that uses reduced number of variables. Although LIBS technique proved to be efficient for predicting fertility attributes in tropical soils, further research is encouraged in order to reduce the amount of sample preparation conducted in this study. •LIBS satisfactorily predicted eight out of nine fertility attributes assessed•iSPA-PLS proved to be the best method for predicting fertility attributes with LIBS•The multivariate models tested are less matrix-matching dependent•Multivariate models using LIBS data allowed satisfactory predictions of ex-P