"Classificazione Automatica delle Specie Vegetali per l'Analisi dei Cambiamenti Climatici Mediante Dati Multispettrali da UAV. Il Caso Studio di Due Aree Lacustri nelle Alpi Marittime." = Automatic classification of vegetation species for climate change analysis by mean of UAV multispectral data. The case study of two lakes areas in the Marittime Alps.

Alpine ecosystems are very responsive to climate change and environmental disturbances, and therefore precise and efficient vegetation surveillance is a key part of ecological research and conservation planning. This thesis describes a multi-source, multi-scale automated land cover and vegetation classification methodology that combines high-resolution UAV-derived multispectral images with medium-resolution Sentinel-2 satellite data. The two target alpine lakes, Lago Vej del Bouc and Lago Brocan, are in the Maritime Alps, and they are used as vegetation change indicators for high-altitude landscapes. Photogrammetric surveys by UAV in 2024 generated ultra-high-resolution five-band orthophotos in spectral bands RGB, Red, Green, RedEdge, and NIR and were processed in Agisoft Metashape using a band-separate workflow for better spectral matching. Sentinel-2 data for the period from 2017-2024 were, in contrast, analyzed in Google Earth Engine for derivation of NDVI, NDWI, EVI, and SAVI indexes for all seasonal time intervals (June, August, and September), allowing trend as well as large-scale classification analysis. Object-Based Image Analysis (OBIA) was carried out on eCognition with multi-resolution segmentation, index calculation, and supervised classification. The following machine learning algorithms—Bayesian, Random Forest, SVM, KNN, and Random Tree—were applied in two classification stages: a primary, general land cover classification (water, soil/rock, vegetation) and, in a subsequent stage, a vegetation type classification with ground-truth data collected in the field. Bayes classifier outperformed more complex models for all classifications, obtaining highest spatial coherence as well as highest classification accuracy for both levels of classification. UAV images proved more helpful for high-resolution vegetation mapping, but satellite data provided useful temporal continuity in addition to large-area context. The outcome confirms the efficacy of a dual-scale, object-oriented scheme for simulating highly complex Alpine landscapes. The resulting scheme forms a reusable template for later environmental monitoring and confirms the value of algorithm selection in reaction to complexity in landscapes and data attributes.