This thesis investigates the use of very-high-resolution PlanetScope satellite imagery combined with the XGBoost machine learning algorithm for snow-cover area (SCA) mapping in the Italian Alps. Five PlanetScope images acquired over the Lombardy region between November 2022 and April 2023 were analysed. A six-class classification scheme was defined to distinguish open snow, forest snow, shadow open snow, shadow forest snow, no snow, and shadow no snow, allowing the effects of forest canopy and terrain shadow on snow detection to be assessed explicitly. Two classification strategies were compared: a scene-based approach, and a global approach. Classification performance was evaluated using independent test samples, and agreement comparisons with the PlanetScope UDM2 product. Additional binary snow/no-snow comparisons were carried out, and terrain shadow was further assessed using hillshade-derived shadow maps. Comparison with UDM2 showed high agreement for Snow and Shadow categories, but lower agreement for Clear/no snow. Excluding Forest snow and Shadow forest snow improved agreement metrics across all dates, indicating that forested snow conditions were a major source of disagreement between the XGBoost classifications and UDM2, indicating that forested terrain is the dominant source of uncertainty in snow-cover mapping with PlanetScope imagery. The results demonstrate that XGBoost combined with topographic auxiliary variables can effectively map snow cover in complex alpine terrain, while highlighting the limitations imposed by the absence of shortwave-infrared bands in PlanetScope products.
Questa tesi analizza l’utilizzo di immagini satellitari PlanetScope ad altissima risoluzione spaziale, combinate con l’algoritmo di apprendimento automatico XGBoost, per la mappatura dell’area coperta da neve (snow-cover area, SCA) nelle Alpi italiane. Sono state analizzate cinque immagini PlanetScope acquisite sulla regione Lombardia tra novembre 2022 e aprile 2023. È stato definito uno schema di classificazione a sei classi per distinguere neve in aree aperte, neve in aree forestali, neve in ombra in aree aperte, neve in ombra in aree forestali, assenza di neve e assenza di neve in ombra, permettendo di valutare esplicitamente gli effetti della copertura forestale e dell’ombra topografica sul rilevamento della neve. Sono state confrontate due strategie di classificazione: un approccio scene-based, in cui un modello è stato addestrato separatamente per ciascuna data di acquisizione, e un approccio globale, basato su un unico modello addestrato utilizzando campioni provenienti da tutte le date. Le prestazioni della classificazione sono state valutate mediante campioni di test indipendenti e mappe di accordo con il prodotto PlanetScope UDM2 per caratterizzare la distribuzione spaziale degli errori. Sono stati inoltre eseguiti confronti binari neve/non neve, mentre l’ombra del terreno è stata analizzata mediante mappe di ombreggiamento derivate da hillshade. Il confronto con UDM2 ha mostrato un elevato accordo per le categorie Neve e Ombra, ma un accordo più basso per la categoria Clear/no snow. L’esclusione delle classi Forest snow e Shadow forest snow ha migliorato le metriche di accordo per tutte le date, indicando che le condizioni di neve in aree forestali rappresentano una delle principali fonti di disaccordo tra le classificazioni XGBoost e UDM2. I risultati dimostrano che XGBoost, combinato con variabili topografiche ausiliarie, può mappare efficacemente la copertura nevosa in terreno alpino complesso, evidenziando al tempo stesso le limitazioni dovute all’assenza di bande nelle lunghezze d'onda dell'infrarosso medio (SWIR-Shortwave Infrared) che risultano particolarmente utili per discriminare la presenza di neve nei prodotti PlanetScope.
Snow cover mapping in Alpine regions using PlanetScope images and machine learning algorithms
ASADIAGHBOLAGHI, HANANEH
2025/2026
Abstract
This thesis investigates the use of very-high-resolution PlanetScope satellite imagery combined with the XGBoost machine learning algorithm for snow-cover area (SCA) mapping in the Italian Alps. Five PlanetScope images acquired over the Lombardy region between November 2022 and April 2023 were analysed. A six-class classification scheme was defined to distinguish open snow, forest snow, shadow open snow, shadow forest snow, no snow, and shadow no snow, allowing the effects of forest canopy and terrain shadow on snow detection to be assessed explicitly. Two classification strategies were compared: a scene-based approach, and a global approach. Classification performance was evaluated using independent test samples, and agreement comparisons with the PlanetScope UDM2 product. Additional binary snow/no-snow comparisons were carried out, and terrain shadow was further assessed using hillshade-derived shadow maps. Comparison with UDM2 showed high agreement for Snow and Shadow categories, but lower agreement for Clear/no snow. Excluding Forest snow and Shadow forest snow improved agreement metrics across all dates, indicating that forested snow conditions were a major source of disagreement between the XGBoost classifications and UDM2, indicating that forested terrain is the dominant source of uncertainty in snow-cover mapping with PlanetScope imagery. The results demonstrate that XGBoost combined with topographic auxiliary variables can effectively map snow cover in complex alpine terrain, while highlighting the limitations imposed by the absence of shortwave-infrared bands in PlanetScope products.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_07_AsadiAghbolaghi.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Thesis text
Dimensione
6.8 MB
Formato
Adobe PDF
|
6.8 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/260679