network for on-site quantitative analysis of soils using laser induced breakdown spectroscopy. Spectrochimica Acta Part B: Atomic Spectroscopy, Elsevier, 2013Elsevier, , 78-79, pp.51-57. 10.1016Elsevier, /j.sab.2012 Artificial neural network for on-site quantitative analysis of soils using laser induced breakdown spectroscopy Nowadays, due to environmental concerns, fast on site quantitative analyses of soils are required. Laser in duced breakdown spectroscopy is a serious candidate to address this challenge and is especially well suited for multi elemental analysis of heavy metals. However, saturation and matrix effects prevent from a simple treatment of the LIBS data, namely through a regular calibration curve. This paper details the limits of this ap proach and consequently emphasizes the advantage of using artificial neural networks well suited for non linear and multi variate calibration. This advanced method of data analysis is evaluated in the case of real soil samples and on site LIBS measurements. The selection of the LIBS data as input data of the network is particularly detailed and finally, resulting errors of prediction lower than 20% for aluminum, calcium, cop per and iron demonstrate the good efficiency of the artificial neural networks for on site quantitative LIBS of soils.

International audienceArtificial neural networks were applied to process data from on-site LIBS analysis of soil samples. A first artificial neural network allowed retrieving the relative amounts of silicate, calcareous and ores matrices into soils. As a consequence, each soil sample was correctly located inside the ternary diagram characterized by these three matrices, as verified by ICP-AES. Then a series of artificial neural networks were applied to quantify lead into soil samples. More precisely, two models were designed for classification purpose according to both the type of matrix and the range of lead concentrations. Then, three quantitative models were locally applied to three data subsets. This complete approach allowed reaching a relative error of prediction close to 20%, considered as satisfying in the case of on-site analysis

No abstract

No abstract