Abstract
Mountain grasslands are increasingly threatened by climate change, land abandonment, and overexploitation. Remote sensing is a valuable tool for monitoring these changes through vegetation mapping. However, challenges such as frequent cloud cover, short growing seasons, and limited field data can reduce the accuracy of results. In this study, we evaluated the effectiveness of different remote sensing data for classifying mountain grasslands in the Sciliar-Catinaccio Nature Park, Italy. We compared classification results using a hyperspectral PRISMA image (Sept 29, 2023), multispectral data from a single-date Sentinel-2 image (Sept 25, 2023), and Spectral-Temporal Metrics (STM) derived from a Sentinel-2 time series from 2021 to 2023. Additionally, we assessed the impact on accuracy of combining optical datasets with Synthetic Aperture Radar (SAR) data, including a time series of 2023 Sentinel-1 backscatter and coherence metrics. Using the Recursive Feature Elimination algorithm (RFE), we selected the most relevant features for classification and applied both Random Forest (RF) and Support Vector Machines (SVM) classifiers. SVM outperformed RF, performing better with the limited training data available. SAR data did not significantly enhance classification and was therefore excluded by the RFE algorithm. PRISMA-based classification achieved up to 74 % accuracy, while single-date Sentinel-2 imagery reached 52 %. The use of STM improved classification performance, yielding an overall accuracy of 77 %. The highest accuracy (87 %) was achieved by combining PRISMA and STM features. These findings suggest that while individual optical datasets may not provide optimal classification accuracy, integrating data from multiple optical sensors significantly enhances the mapping of mountain grasslands.