Publicaciones Científicas

Satellite embeddings—compact numerical representations derived from deep learning models applied to satellite imagery—offer a new paradigm for environmental monitoring that bypasses traditional spectral index selection. This study demonstrates that 64-dimensional embeddings from a Vision Transformer (ViT) model, accessed through Google Earth Engine's public ImageCollection, can detect and classify volcanic surface changes without supervised training or manual feature engineering.

Applied to Nevados de Chillán and Llaima volcanoes in Chile (2017–2023), our unsupervised approach identified distinct surface classes through K-means clustering of embeddings, tracked temporal evolution through embedding distance metrics and UMAP trajectory visualization, and detected both dramatic post-eruption changes and subtle ongoing surface modifications.

At Nevados de Chillán, the method revealed four distinct surface classes and captured the progressive stabilization following the 2016–present eruption cycle. At Llaima, three surface classes were identified with notably smaller temporal changes, consistent with the volcano's current quiescent state. The embedding-based approach achieved high cluster separability (silhouette scores of 0.52–0.59) and produced physically interpretable results consistent with known volcanic activity.

These findings suggest that satellite embeddings could complement traditional remote sensing approaches for operational volcano monitoring, particularly for detecting gradual surface changes that may precede volcanic unrest.