A hierarchical approach to classify rock units using airborne geophysical and geological field data

Supervised image classification is based on assembling statistics between ground observations and remotely sensed measurements. If supervised image classification is applied within the context of a particular theme (e.g., vegetation, soil, lithology, land use), one is often confronted with extracting the statistical correlations from a hierarchy of classes (e.g., a taxonomy). The spatial arrangement of the classes often reflects the hierarchical structure of the taxonomy because the inference, from more specific to less specific classes, is largely based on a hierarchically structured generalization over a range of observation and mapping scales. However, in practice, supervised image classification often appears to be based on a pragmatic approach, a priori, assigning classes into a linear schema derived from various levels or from a subset of the taxonomy. The classification performance is, as a result, often assessed with respect to arbitrary class schemas that only partly correspond to the schemas obtained by field surveys. Clearly, to gain more insight to the extent that supervised image classification can support field surveys, sampling procedures are required that respect the inherent hierarchy of the taxonomy used, within a particular geoscience discipline.

In this paper, a hierarchical approach to supervised image classification is presented and applied to the prediction of lithological units of the exposed Canadian Shield of northeastern Alberta. Samples were extracted from a geological field database to establish the statistical relationships between outcrops classified through a four-level lithostratigraphic class schema and airborne magnetic and gamma-ray spectrometry data. The number of classes in the lithostratigraphic schema ranges from lithotectonic assemblages (n = 2) for the highest level and to rock units (n = 14) for the lowest level. The results of the classification experiments suggest that the performance of supervised image classification can be improved if the estimation of prior probabilities, at a more detailed level of the taxonomy, is conditioned by spatial patterns at more general levels.