Remote sensing-based approach for monitoring vegetation dynamics under climate and land use changes

Alpine environments are increasingly threatened by a combination of anthropogenic pressures or, on the opposite, abandonment, and climate change, resulting in habitat transformations. Monitoring these dynamics is essential to support biodiversity conservation and evidence-based management strategies. In this context, this thesis aims to demonstrate how the integration of satellite remote sensing, geospatial analytics, and machine learning can contribute to this objective by detecting and quantifying long-term trends and their underlying drivers, thereby providing a robust framework for generating spatially and temporally explicit information. Three complementary research topics were addressed. First, a novel workflow was implemented to retrieve daily high-resolution (20 m) snow cover maps by combining MODIS and Sentinel-2 imagery. Three random forest-based approaches were compared, differing in input data and model structure. The automated workflow, validated with both Sentinel-2 and in situ meteorological data, achieved high accuracy (92.6-93.7%) and successfully captured seasonal snow phenology, offering a powerful tool for monitoring snow-driven ecosystem processes (Richiardi et al., 2023). Second, a cumulative spatial and temporal assessment of anthropogenic impacts, including tourism infrastructure, overgrazing, helicopter flights, and hydroelectric powerplants, was developed using a normalized, weighted GIS-based methodology. The resulting risk map, derived by combining impact intensity and ecological vulnerability, reveals that while most of the park falls under low or no risk, specific sensitive habitats such as lentic waters and mountain pine forests show substantial exposure to pressure (Richiardi et al., 2023). Third, a hierarchical method for long-term habitat classification was developed to map changes from 1985 to 2023 using Landsat time series. The approach relies on a single pre-existing habitat map, a digital elevation model, and seasonal composite imagery. A Z-statistics-based training sample selection ensured consistency and ecological relevance. Results indicate that 88% of the area remained stable over time, while trends include a loss of grasslands (−10 ha yr⁻¹) and an increase in shrubs (+10 ha yr-1). The approach proved robust, scalable, and suited to protected areas where detailed long-term field data are scarce (Richiardi et al., 2025).