In this paper, we demonstrate the importance of variable selection on the prediction ability of LIBS quantitative partial least squares (PLS) models. The spectral lines of potassium at 766.49 nm and 769.90 nm were considered in the framework of an agricultural soils analysis. Univariate models demonstrating very poor correlation between the peak areas of the potassium lines and the related concentration values, a series of PLS models allowed to significantly improve the prediction ability compared to the univariate approach. This improvement was due to advanced variable selection, achieved through the use of two output data provided after PLS calculation, namely the Variable Importance in Projection (VIP) and the Coefficients graph. In this demonstration, the gain was significant because the two spectral lines of potassium at 766.49 nm and 769.90 nm exhibited unusual profiles. Indeed, including in a PLS model only the variables related to the edges of these lines allowed a significant improvement of its predictive ability (Q2 = 0.84, RMSE = 1.49 g/kg) compared to another PLS model only including the variables related to the central parts of these lines (Q2 = 0.78, RMSE = 1.59 g/kg).< Leer menos