SATELLITE IMAGE TIME SERIES CHANGE VECTOR ANALYSIS FOR LAND COVER CHANGE DETECTION

The exponential increase in the temporal availability of Satellite Image Time Series (SITS), driven by missions such as Sentinel-2, has fundamentally transformed the capability to monitor Earth’s surface dynamics. However, the exploitation of this data density is currently hindered by a methodological trade-off: while Deep Learning offers high predictive performance, it relies on prohibitive amounts of labeled data and lacks geometric transparency. Conversely, traditional physical methods like Change Vector Analysis (CVA) offer interpretability but are mostly constrained to bi-temporal comparisons, failing to capture the continuous evolution of complex landscapes. This thesis addresses this gap by proposing and formalizing the Time Series Change Vector Analysis (TSCVA) framework. The primary objective is to generalize the geometric principles of vector analysis into the high-dimensional temporal domain, thereby restoring physical interpretability to unsupervised change characterization. Methodologically, the research is structured into two phases. First, the thesis establishes the theoretical foundations of the spectral–temporal vector space. It demonstrates that by treating the time series within a joint spectro-temporal domain, change can be mathematically decomposed into Magnitude (intensity of change) and Direction (type of change). A semi-supervised implementation validates that these high-dimensional directional variables carry consistent semantic meaning, allowing for the discrimination of multi-class transitions with minimal expert guidance. Second, the thesis extends this framework into a fully automated, unsupervised pipeline. To solve the challenge of reference definition without prior knowledge, a novel clustering-based strategy is introduced. This method automatically extracts stable spectro-temporal centroids from the data structure, serving as the mathematical origin for vector calculation. Integrated with spatial-contextual constraints to mitigate atmospheric noise and registration errors, the pipeline effectively decouples structural land cover changes (e.g., deforestation, burnt areas, urbanization) from natural phenological variability. Experimental validation across diverse environmental conditions demonstrates that TSCVA outperforms standard baselines in scenarios where labeled data is scarce. By providing a physically meaningful, reproducible, and data-driven standard for change characterization, this work establishes a robust foundation for the next generation of global, unsupervised environmental monitoring systems.