Habitat loss is the main threat to biodiversity at a global level, making habitat mapping an essential tool for the management of protected areas and for the conservation and monitoring, in line with Directive 92/43/EEC. Traditional mapping methods are resource-intensive, while remote sensing approaches depend on the availability of ground truth datasets. In this context, this study presents a novel framework for time series habitat classification. The approach leverages a single pre-existing habitat cartography and a limited set of ancillary data to derive a retrospective training dataset. The method was applied to analyse 39 years (1985–2023) of habitat and land cover changes in Gran Paradiso National Park (NW Italy). Annual seasonal composite images were generated for the growing and senescence seasons using an enhanced Best Available Pixel approach. Annually derived training datasets were used to classify hierarchically land cover and habitats via ensemble random forest models. Validation against high-resolution maps demonstrated the robustness of the approach. The method allows for long-term habitat monitoring even in data-sparse environments. The results reveal high stability of land cover (88% of the area) and significant trends in some vegetation types, including a decline in grasslands (− 10 ha year⁻¹) and the expansion of shrublands (+ 10 ha year⁻¹). The method proved to be reliable for large patches, less so for ecotones. Future research should explore its application across different landscapes. This work underscores the potential of remote sensing for long-term habitat monitoring, providing a cost-effective solution to support biodiversity preservation efforts. © 2025 Elsevier B.V., All rights reserved.

Cited by: 0; All Open Access; Hybrid Gold Open Access