Teacher
Vasil Yordanov (vasil.yordanov@polimi.it)
Tutors
Polychronis Kolokousis (polychronis.kolokousis@polimi.it)
Research area
Geoinformatics
Keyword
Geospatial foundation models, Spatial embeddings, Down-sampling, Multi-scale representation
Technologies
Languages: Python, Jupyter Notebook.
Description
This project proposes an investigation into aggregation strategies for AlphaEarth Foundations (AEF), with a focus on evaluating the effect of different down-sampling scales and methods on the resulting representation quality. The study shall involve a systematic comparison of embedding representations obtained at varying spatial resolutions, analysing how the choice of down-sampling approach influences the geometry and structure of the embedding space. To this end, pairwise similarity distributions should be computed across scales and subsequently normalised to enable a statistically grounded comparison. The goal is to assess whether and to what extent the proposed aggregation methods preserve or distort the discriminative properties of the AEF embeddings, ultimately providing guidelines for selecting appropriate down-sampling configurations in practical remote sensing and geospatial learning pipelines.