Respecting Hierarchically Structured Taxonomies in Supervised Image Classification: a Geological Case Study from the Western Canadian Shield

Supervised image classification is based on assembling statistics between site-specific 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 hierarchically arranged network of taxonomic classes spatially abstracted and hierarchically generalized over a range of mapping scales. In practice, however, supervised image classification often appears to be based on a pragmatic approach, a priori categorizing the samples into classes from various levels or from a subset of the hierarchic network. Such approaches are suspect, since the sampling often appears to be biased towards maximizing the discrimination potential of the multivariate data set at cost of representing the categories identified by direct ground observation. The classification performance is, as a result, often assessed within the context of arbitrarily defined class schemas that only partly correspond to the schemas obtained by field surveys. Clearly, to gain more insight in how supervised classifiers are behaving with respect to ground observations, sampling procedures are required that respect the hierarchy of the taxonomy obtained in ground surveys. Herein we report the results of classification experiment applied to gamma-ray spectrometry and aeromagnetic data where samples are extracted at ground stations for each level of a four-level hierarchically arranged class network of bedrock lithology. The number of classes in this network ranges from n = 2 for the highest level and to n = 14 for its lowest. A number of classification experiments suggest that classification performance can be improved if the estimation of prior probabilities at a more detailed level in the taxonomy is conditioned by spatial patterns at more general levels in the taxonomy. This improvement in performance may even apply when such patterns are obtained by classification of the same data and sample set.