Application of a series of artificial neural networks to on-site quantitative analysis of lead into real soil samples by laser induced breakdown spectroscopy

TLDR: Artificial neural networks were applied to process data from on-site LIBS analysis of soil samples as discussed by the authors, which allowed retrieving the relative amounts of silicate, calcareous and ores matrices into soils.

About: This article is published in Spectrochimica Acta Part B: Atomic Spectroscopy.The article was published on 2014-07-01 and is currently open access. It has received 61 citations till now. The article focuses on the topics: Soil test & Artificial neural network.

Figures

Table 1 Wavelength (nm) of the 35 spectral lines selected for the LIBS analyses.

Table 6 Performance of the three ANN models calculated after Y-randomization. The values reported in the table are average values calculated after 25 repetitions (see text for details). The letters c, v, and t stand for calibration, validation and test sets, respectively.

Table 4 Performance of the two quantitative ANN models applied to the LIBS data in the case of quantitative analysis of soil samples; ANN1 for lead concentrations above 10 000 ppm and ANN2 for lead concentrations below this value.

Table 5 Performance of the two quantitative ANN models applied to the LIBS data in the case of quantitative analysis of soil samples and for lead concentrations above 10 000 ppm. The model ANN3was optimized for analyzing samples characterized by amatrix b80% silicate and the model ANN4 was optimized for samples with N80% silicate.

Table 2 Performance of the ANN model optimized to provide the relative values V1, V2 and V3 described in the text and related to the three poles of the ternary diagrams (cf. Figs. 1 and 2). C, V and T stand for calibration, validation and test set, respectively.

Table 3 Performance of the ANN model applied to the whole set of LIBS data in the case of quantitative analysis of soil samples.

Frequently asked questions

Q1. What are the contributions in "Application of a series of artificial neural networks to on-site quantitative analysis of lead into real soil samples by laser induced breakdown spectroscopy" ?

In this paper, a special focus on the analysis of lead is presented, where the authors use ANN to predict the concentrations of major elements such as calcium, aluminum and iron and also those of trace elements as copper. 

Q2. What future works have the authors mentioned in the paper "Application of a series of artificial neural networks to on-site quantitative analysis of lead into real soil samples by laser induced breakdown spectroscopy" ?

Further work should be dedicated to build a growing database of soils in order to continue to enhance the performance of the ANN models for quantitative LIBS. The strategy consists in building as many ANN models as necessary in order to be able to analyze in the future any soil sample, whatever its matrix. 

Q3. How many locations of the laser spot were used to calculate the average spectrum?

side experiments allowed verifying that averaging over 25 locations of the laser spot at the sample surface was sufficient to correctly take into account the sample's heterogeneity. 

Q4. How many laser shots were used to optimize to signal-to-noise ratio?

To optimize to signal-to-noise ratio, it was decided that each LIBS spectrumwould be the result of 25 laser shots accumulated at the same point of the sample, with a gate delay of 300 ns and a gate width of 3 μs. 

Q5. What type of ores should be associated with high concentrations of Pb?

High values of concentration for Pb should be related to the natural ore of Galena (PbS) while the presence of Zn could be related to two types of natural ores, namely sphalerite (ZnS) and calamine (ZnCO). 

Q6. How many neurons were found in the hidden layer?

Based on themethod of external validation and using data fromboth the calibration and the validation sets, the optimized parameters for the ANN were found to be: number of neurons into the hidden layer: 4, learning rate: 0.01, momentum: 0.1, and number of iterations: 19 000. 

Q7. How many samples were used in the previous study?

The 117 soil samples providedby the 3 campaigns of on-site LIBS measurements were split in the sameway as for the previous study, namely 76 samples into the calibration set, 21 into the validation set and 20 into the test set. 

Q8. What was the consequence of the Y-randomization procedure?

As expected, the consequence of the Y-randomization procedure was to drastically decrease the predicting ability of the ANNmodels. 

Q9. What methods have been used to classify soil samples?

More precisely, several multivariate methods such as PCA, SIMCA, LDA and PLS-DA have been applied to classify soil or geomaterial samples [16–21]. 

Q10. What is the potential of laser-induced breakdown spectroscopy?

Laser-induced breakdown spectroscopy is recognized to have high potential for geochemical applications since this technique is able to achieve rapid and multi-elemental on-site analysis with very little sample preparation [1–5]. 

Q11. How many samples were split into the calibration set?

The 117 original samples were split into three data subsets: 76 into the calibration set, 21 into the validation set and 20 into the test set. 

Q12. What is the importance of taking into account spectral lines from the matrix in addition to those?

In thispast work, the authors highlighted the importance of taking into account spectral lines from the matrix in addition to those of the analyte as input data of the ANN in order to improve the prediction ability of the model [28]. 

Q13. How many input values were retrieved by the ANN model?

The three output values calculated by the ANN model ranged between 0 and 1 and thus could be directly interpreted as percentage values without any additional treatment. 

Q14. How many micrograms of matter were analyzed by LIBS?

And despite of the very small amount of matter analyzed by LIBS for each soil sample, typically in the range of hundreds of micrograms, relative error of prediction as low as 20 % was obtained. 

Q15. How did the authors decide to classify the samples?

The authors decided to check the ability of ANN to classify the samples into two classes by setting a threshold value of 10 000 ppm for the lead concentration.