Source apportionment of soil samples by the combination of two neural networks based on computer-controlled scanning electron microscopy

The apportionment of ambient aerosol mass to different sources of airborne soil is a difficult problem because of the similarity of the chemical compo sition of crustal sources. However, additional information can be obtained using individual particle analysis. A novel approach based on the combinati on of two neural networks, the adaptive resonance theory-based neural netwo rk (ART-2a) and the back-propagation (BP) neural network with electron micr oscopy data, has been developed to apportion the mass contributions of the crustal sources to ambient particle samples. The crustal source samples wer e analyzed using computer-controlled scanning electron microscopy (CCSEM). CCSEM provides elemental compositions and size parameters for individual pa rticles as well as estimates of the shape and density from which the volume and mass of each particle can be estimated. The ART-2a neural network was first used to partition particles into homogeneous classes based on the ele mental composition data. After the different particle type classes were pro duced by ART-2a, their mass fractions were calculated. In this way, the sou rce profiles for the crustal dust sources can be obtained in terms of the m ass fractions for different particle types. Then the BP neural network was applied to build the model between the mass fractions of different particle types and the mass contributions. Using the three physical source samples prepared for this study, artificial ambient samples were generated by rando mly mixing particles from the three source samples. These samples were then used to examine the proposed method. Satisfactory predictions for the mass contributions of the three sources to the ambient samples have been obtain ed, indicating the proposed method is a promising tool for the source appor tionment of chemically similar soil samples.