Classification of airborne particles by analytical scanning electron microscopy imaging and a modified Kohonen neural network (3MAP)

A projection of Primʹs minimal spanning tree into a Kohonen neural network provided a method for visual unsupervised chemical pattern recognition. Digitized scanning electron microscopy images of airborne particles from distinct sources of air pollution were mathematically transformed to multivariate descriptors x characterizing the shape of each particle. These m-dimensional descriptors x(m ⪢ 3) were classified using a visually interpretable Kohonen neural network. The obtained two-dimensional projection preserved the topology of the original m-dimensional space, but a part of information about correct distances between the descriptors has been lost. This loss has partly been compensated for by computation of Primʹs Minimal Spanning Tree between the loaded neurons of the trained neural network. This combination method, called 3MAP, allowed a quantitative visualization of clusters but also trends and transitions between several types of spherical and non-spherical, hexagonal, octagonal, ellipsoidal, triangular, rectangular and circular shapes of airborne particles. Combining then shape information with chemical-analytical composition of a particle in a common extended descriptor x allowed the visual identification of particles that were emitted by distinct sources of air pollution. Compared with principal component analysis and multidimensional scaling, the new 3MAP combination method provided an easier interpretable visualization of the m-dimensional variables space.