Environmental Characterization and Recognition from Raw GNSS Data Collected with a Low-Cost Receiver Caratterizzazione e Riconoscimento Ambientale da Dati GNSS Grezzi Raccolti con un Ricevitore a Basso Costo Caratterizzazione e Riconoscimento Ambientale da Dati GNSS Grezzi Raccolti con un Ricevitore a Basso Costo

The diffusion of multi-frequency, multi-constellation GNSS chipsets in modern smart- phones has made high-accuracy satellite positioning increasingly accessible. How- ever, the limited antenna quality and hardware constraints of low-cost receivers make their performance strongly dependent on the surrounding environment: open- sky conditions, foliage, urban canyons, and signal blockage introduce characteristic distortions in the received signals that degrade positioning accuracy in different and often unpredictable ways. This thesis presents a two-part study conducted during an internship at Topcon Positioning Italy. The first part evaluates the RTK positioning performance of a Huawei Mate 40 Pro smartphone, used with both an external antenna setup and its internal antenna, through a series of static and kinematic tests. The results confirm that centimeter-level accuracy is achievable even with low-cost hardware, although reliability decreases under challenging signal conditions. The second part addresses the automatic recognition of the receiver’s surround- ing environment from raw GNSS measurements. A complete Machine Learning pipeline was developed, starting from the Google Smartphone Decimeter Challenge 2023/2024 dataset. Eight epoch-level features were engineered from raw GNSS ob- servations, including satellite elevation distribution, PDOP, number of tracked satel- lites, average SNR, and SNR temporal variation. Since no ground-truth environ- ment labels were available, an unsupervised labeling procedure was designed using HDBSCAN clustering applied to the standardized, PCA-reduced feature space; the resulting labels were validated geographically through Google Earth visualization. An XGBoost multiclass classifier was then trained to recognize three environment classes – open-sky, partial shadowing, and blockage – with hyperparameters tuned via RandomizedSearchCV. To the best of the author’s knowledge, this is the first ap- plication of XGBoost to epoch-level GNSS environment recognition using automat- ically generated labels. The model demonstrated promising results on independent kinematic test data, with visible improvements after hyperparameter tuning.