Capability of feed-forward neural networks for a chemical evaluation of sediments with diffuse reflectance spectroscopy

Diffuse reflectance spectroscopy (0.4–2.5 μm) is evaluated as fast and non-destructive method for the analysis of sediments, characterised by a wide range of mineral constituents. Combined with feed-forward artificial neural networks (ANNs) this technique is used to estimate quantitatively the chemical composition from the sediments based on a supervised training with one model. The examined characteristics include contents of inorganic carbon, Fe, S, Al, Si, Ca, K, Mg and calcite. The efficiency of several learning algorithms (Backpropagation, Quickprop, Resilient propagation (Rprop), Cascade Correlation (CC)) is investigated. All learning algorithms perform well using principal component (PC) scores of the first derivative spectra as input for the supervised training. ANNs trained with Quickprop and Rprop produced most accurate estimations of the chemical characteristics and the performance was better than for standard multivariate statistical tools (stepwise multiple linear regression (SMLR), principal component analysis (PCA)). An interpretation of the results is given by a detailed consideration of the correlation structure among the chemical constituents.