The increasing availability of high-resolution satellite imagery has made Earth Observation a rapidly evolving field. Importantly, satellite data are inherently non-stationary, with distributions evolving over time due to land-cover and environmental changes (temporal domain shift) and varying across geographic regions (spatial domain shift), which makes continuously collected satellite data streams particularly challenging to analyze. Although traditional learning models perform well in satellite image classification, they are typically trained offline on static datasets, limiting their ability to adapt to evolving data. In this context, Streaming Machine Learning (SML) addresses this by enabling incremental learning, real-time updates, and adaptation to distributional shifts. This thesis proposes a framework that integrates SML with the foundation model DINOv3 for feature extraction from satellite image time series. In this framework, dense self-supervised representations from DINOv3 are combined with streaming learning to achieve robust and incremental performance. The pipeline is organized into multiple processing stages to enhance learning effectiveness: dimensionality reduction and normalization improve computational efficiency, hyperparameter tuning refines model memory and adaptation behavior, and prequential evaluation enables continuous updates, allowing the model to adapt to domain shifts and improve performance. The framework is evaluated on the DynamicEarthNet dataset, which includes two years of satellite observations across multiple Areas of Interest (AoIs), naturally encompassing both temporal and spatial domain shifts. Experimental results demonstrate that combining self-supervised representations with streaming learning enables strong generalization. Overall, the proposed pipeline offers an efficient solution for near-real-time land-cover classification of satellite imagery under evolving conditions.
La crescente disponibilità di immagini satellitari ad alta risoluzione ha reso l’Osservazione della Terra un campo in rapida evoluzione. I dati satellitari sono intrinsecamente non stazionari: le distribuzioni cambiano nel tempo a causa di variazioni della copertura del suolo e fattori ambientali (temporal domain shift) e differiscono tra regioni geografiche (spatial domain shift), rendendo particolarmente complessa l’analisi dei flussi continui di dati satellitari. Sebbene i modelli tradizionali siano efficaci nella classificazione della immagini satellitari, sono generalmente addestrati offline su dataset statici, limitando la loro capacità di adattarsi ai dati in evoluzione. In questo contesto, lo Streaming Machine Learning (SML) permette l'apprendimento incrementale, aggiornamenti in tempo reale e adattamento ai cambiamenti nella distribuzione dei dati. Questa tesi propone un framework che integra SML con il modello foundation DINOv3 per l’estrazione di feature da serie temporali di immagini satellitari. Le rappresentazioni dense auto-supervisionate di DINOv3 sono combinate con l'apprendimento streaming per ottenere prestazioni robuste e incrementali. La pipeline è organizzata in più fasi: riduzione dimensionale e normalizzazione migliorano l'efficienza computazionale, l’ottimizzazione degli iperparametri affina memoria e adattamento del modello, e la valutazione prequenziale consente aggiornamenti continui, permettendo al modello di adattarsi ai domain shift e migliorare le performance. Il framework è valutato sul dataset DynamicEarthNet, che include due anni di osservazioni satellitari su più aree, comprendendo naturalmente sia shift temporali che spaziali. I risultati dimostrano che la combinazione di rappresentazioni auto-supervisionate con l’apprendimento streaming garantisce una forte generalizzazione. Nel complesso, la pipeline proposta offre una soluzione efficiente per la classificazione del suolo da immagini satellitari in scenari reali in continua evoluzione.
Satellite image classification under domain shifts using self-supervised learning and streaming machine learning
Nguyen Ba, Chiara Thien Thao
2024/2025
Abstract
The increasing availability of high-resolution satellite imagery has made Earth Observation a rapidly evolving field. Importantly, satellite data are inherently non-stationary, with distributions evolving over time due to land-cover and environmental changes (temporal domain shift) and varying across geographic regions (spatial domain shift), which makes continuously collected satellite data streams particularly challenging to analyze. Although traditional learning models perform well in satellite image classification, they are typically trained offline on static datasets, limiting their ability to adapt to evolving data. In this context, Streaming Machine Learning (SML) addresses this by enabling incremental learning, real-time updates, and adaptation to distributional shifts. This thesis proposes a framework that integrates SML with the foundation model DINOv3 for feature extraction from satellite image time series. In this framework, dense self-supervised representations from DINOv3 are combined with streaming learning to achieve robust and incremental performance. The pipeline is organized into multiple processing stages to enhance learning effectiveness: dimensionality reduction and normalization improve computational efficiency, hyperparameter tuning refines model memory and adaptation behavior, and prequential evaluation enables continuous updates, allowing the model to adapt to domain shifts and improve performance. The framework is evaluated on the DynamicEarthNet dataset, which includes two years of satellite observations across multiple Areas of Interest (AoIs), naturally encompassing both temporal and spatial domain shifts. Experimental results demonstrate that combining self-supervised representations with streaming learning enables strong generalization. Overall, the proposed pipeline offers an efficient solution for near-real-time land-cover classification of satellite imagery under evolving conditions.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_03_NguyenBa_Executive_Summary.pdf
accessibile in internet per tutti a partire dal 02/03/2027
Descrizione: executive summary
Dimensione
2.27 MB
Formato
Adobe PDF
|
2.27 MB | Adobe PDF | Visualizza/Apri |
|
2026_03_NguyenBa_Tesi.pdf
accessibile in internet per tutti a partire dal 02/03/2027
Descrizione: tesi
Dimensione
15.87 MB
Formato
Adobe PDF
|
15.87 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/10589/251304