Quantifying mineral abundances of complex mixtures by coupling spectral deconvolution of SWIR spectra and regression tree analysis
The following is a summary of the paper by Titia Mulder which is nominated for Best Paper in Pedometrics 2013.
Soil mineralogy is an important indicator for soil formation and parent material characterization. In environmental and geological studies, the characterization (and quantification) of soil mineralogy is typically achieved using X-ray diffraction (XRD). Visible Near Infrared and Shortwave Infrared (VNIR/SWIR) spectroscopy has proven to be an efficient alternative for the determination of various soil properties. In this paper we propose and demonstrate its use for simultaneous quantification of mineral abundances from complex mixtures.
Detection of minerals having absorption features within the 0.350–2.500 µm range have been successfully obtained using linear spectral unmixing techniques. However, these analyses were limited to estimating the main component within a sample having the most distinct absorption feature. Hence, reflectance spectra of mixtures are typically a complex result from the combinations of the spectral characteristics of the constituents. Depending on the composition, the abundance and the spatial arrangement of the minerals, the total reflectance resulting from the scattering of the minerals within the intimate mixture produces positional shifts, changes in intensity, disappearance of absorption features or changes in their shape.
Hence in this work we aimed to quantify mineral abundances using spectral deconvolution (SD) followed by regression tree analysis (RT). SD involves modelling the total reflectance and the inference of absorption components within complex features by fitting (modified) Gaussian curves to the absorption features and absorption components. Next, mineral abundances were predicted by RT using the parameters of the fitted Gaussians as inputs. The approach was demonstrated on a range of prepared samples with known abundances of kaolinite, dioctahedral mica, smectite, calcite and quartz and on a set of field samples from Morocco.
Cross validation showed that the prepared samples of kaolinite, dioctahedral mica, smectite and calcite were predicted with a root mean square error (RMSE) less than 9wt%. For the field samples, the RMSE was less than 8 wt% for calcite, dioctahedral mica and kaolinite abundances. Smectite could not be well predicted, which was attributed to spectral variation of the cations within the dioctahedral layered smectites. Substitution of part of the quartz by chlorite at the prediction phase hardly affected the accuracy of the predicted mineral content; this suggests that the method is robust in handling the omission of minerals during the training phase. The degree of expression of absorption components was different between the field sample and the laboratory mixtures. This demonstrates that the method should be calibrated and trained on local samples. Concluding, our method allows the simultaneous quantification of more than two minerals within a complex mixture and thereby enhances the perspectives of spectral analysis for mineral abundances.
Mulder, V.L., Plötze, M., de Bruin, S., Schaepman, M.E., Mavris, C., Kokaly, R.F., Egli, M., 2013. Quantifying mineral abundances of complex mixtures by coupling spectral deconvolution of SWIR spectra (2.1–2.4 μm) and regression tree analysis. Geoderma 207–208, 279–290.
The Full paper is available as pdf here