The use of remote-sensing techniques in the discrimination of rock and soil classes in northern regions can support a diverse range of activities, such as environmental characterization, mineral exploration and the study of Quaternary paleoenvironments. Although images with low spectral resolution can commonly be used in the mapping of classes possessing distinct spectral properties, hyperspectral images offer greater potential for discrimination of materials characterized by more subtle reflectance properties. In an effort to better constrain the utility of broadband and hyperspectral datasets in high-latitude research, this study investigated the effectiveness of Landsat Thematic Mapper (TM) and EO-1 Hyperion data for discrimination of lithological classes at eastern Melville Island, Nunavut, Canada. TM data were classified using a standard neural-network algorithm, and both TM and Hyperion data were linearly unmixed using ground-truth spectra. TM classification results successfully discriminate between classes over much of the study area, although with incomplete separation between clastic and carbonate materials. TM unmixing results are poor, with useful class separation restricted to vegetation and red-weathered sandstone classes. Hyperion results effectively depict the fractional cover of end members, although the abundance images of several classes contain background abundance values that overestimate surface exposure in some areas. For the study area and surface classes involved, noisy hyperspectral data were found to be of greater utility than higher-fidelity broadband multispectral data in the generation of fractional abundance images for an inclusive set of surface-cover classes.