Urban growth models are useful tools to assess the impact of alternative public policy scenarios. The calibration of such models requires historical land-use data, which is not always available at frequent time intervals. In this research, information on spatio-temporal change of land-cover gradients is extracted from two multidate Landsat images by linear spectral unmixing. For this purpose, the two images were subjected to a relative radiometrical calibration in order to reduce the impact of sensor drift and differences in illumination and atmospheric conditions. The resulting gradient information can be used to derive spatial metrics, which in turn characterise urban form and morphology and may serve as a proxy to land-use data. Because the sub-pixel proportions do not contain implicit information as to whether a particular pixel belongs to the urban fabric, an urban mask was required to make that distinction. This mask was created by first applying an unsupervised classification based on Kohonen self-organising maps, and subsequently enhanced by applying ad-hoc knowledge-based post-classification rules. The results of this study demonstrate that sub-pixel change maps are useful to identify urban growth patterns. In combination with the urban mask and unsupervised classification, sub-pixel gradients will be used in future research to investigate if they provide useful information on urban structure, and can maybe even be used to infer certain types of land use.