Mapping multi-decadal tree cover change from disturbances across Europe using spectral unmixing of Landsat time series

Europe’s forests are under increasing pressure from natural disturbances, while there is growing demand for wood. With disturbances expected to further intensify under climate change, quantifying the impact of disturbances on forest resources has thus become a key challenge for Earth Observation. In particular, there is a strong need for spatially explicit information on how forests change over time, driven by disturbance and post-disturbance recovery. Most existing large-scale assessments derive this information using spectral indices. However, spectral indices tend to saturate in dense canopies while being limited in their ability to capture changes under mixed land cover conditions. In that sense, using estimates of tree cover might provide a clearer signal of forest canopy changes. Building on previous applications of spectral unmixing for mapping forest cover in European forest ecosystems (Mandl et al. 2024, Viana-Soto et al. 2022, Senf et al. 2019), we here present a novel framework for mapping tree cover across all of Europe’s forests. Specifically, we (i) estimate annual tree cover fractions from 1985 to 2024 at 30 m spatial resolution using spectral unmixing of Landsat data, (ii) assess the temporal consistency and accuracy of these estimates across Europe’s forests, and (iii) characterise tree cover loss from disturbance and post-disturbance tree cover gain, thereby distinguishing it from land use changes. As a data basis, we built a consistent Landsat data cube of atmospherically and topographically corrected Landsat surface reflectance data, including cloud and shadow masking, totalling to 363,088 images. Annual gap-free best available pixel composites were generated by selecting high-quality observations closest to 1st of August, minimizing phenological effects and ensure intra-annual consistency. Based on these composites, we developed a multi-year endmember library consisting of pure and temporally stable pixels representing treed and non-treed land cover types (herbaceous, shrubs, bare ground, and shadow). We collected endmember spectra by randomly sampling pure pixels from LUCAS database, providing in-situ land cover information across Europe, and by cross-checking their spectral–temporal stability and cover proportions using high-resolution imagery. To simulate the full range of possible spectral mixtures, we generated synthetic training datasets by linearly combining endmember spectra in known proportions. Lastly, these mixtures and their associated ratios were used to train regression models predicting annual tree cover fractions. Preliminary results indicate that the spectral unmixing framework enables consistent mapping of annual tree cover fractions across Europe, capturing losses associated with disturbance or land use conversion and gradual gains reflecting post-disturbance recovery. By delivering harmonized annual maps of tree cover fractions for Europe, this work advances continental-scale forest monitoring efforts and supports policy frameworks for forest adaptation to climate change.