In this study we developed a methodology aimed at improving the assessment of inter-annual land cover dynamics from hard classified remotely sensed data in heterogeneous and resilient landscapes. The methodology is implemented for the Spanish Natural Park of Sierra de Ancares, where human interference during the last century has resulted in the destruction and fragmentation of the original land cover. We ran supervised classifications, with a maximum likelihood algorithm (Maxlike), on a temporal series of Landsat images (1991–2005), followed by an uncertainty assessment using fuzzy classifications and confusion indices (CIs). This allowed us to show how much (and where) of the resulting maps contained a substantial amount of error, distinguishing data that might be useful to measure land change from data that are not particularly useful when applying a post-classification comparison methodology. In this way, we can detect true changes not skewed by the effects of uncertainty. Even if patterns of change were always coherent amongst years, they were more realistic after reducing uncertainty, in spite of a substantial decrease in the number of available pixels (i.e. unmasked by the method). We then computed land cover dynamics by means of a model specifically designed to determine the frequency of disturbances (mainly fire events) and the vegetation recovery time during the study period. Model outputs showed correlated landscape patterns at a broad scale and provided useful results to explore land cover change from pattern to process.