The Landsat archive offers great potential for monitoring forest cover change, and new approaches moving from categorical towards continuous change products emerge rapidly. Most approaches, however, require vast amounts of high-quality reference data, limiting their applicability across space and time. We here propose the use of a generalized regression-based unmixing approach to overcome this limitation. The unmixing approach relies on temporally generalized machine learning regression models (random forest regression [RFR] and support vector regression [SVR]), which are trained on synthetically mixed data from a multi-year library of pure and hence easy to identify image spectra. We apply the model to three decades of Landsat data, mapping both overall forest cover and broadleaved/coniferous forest cover fractions across space and time. The resulting maps well represented the spatial-temporal patterns of forest (change) in our study region. The SVR model outperformed the RFR model, yielding accuracies of r2 = 0.74/RMSE = 0.18 for the forest cover fraction maps, r2 = 0.50/RMSE = 0.24 for the broadleaved forest cover fraction maps, and r2 = 0.59/RMSE = 0.23 for coniferous forest cover fraction maps, respectively. Highest map errors were found in mature stands, residential areas, and recently disturbed forests. We also found some variability in forest cover fractions for stable forest pixels over time, which were explained by variation in Landsat image acquisition dates. We conclude that regression-based unmixing using synthetically mixed training data from a multi-year spectral library offers an innovative strategy for mapping forest cover fractions and forest types throughout the Landsat archive that likely can be extended to large areas.