Establish robustness of foundation-model embeddings to year-specific data quality

Determine whether geospatial foundation-model embeddings provide robust tree-species classification under variation in cloud cover, observation density, phenological variability, and other year-specific data-quality conditions beyond the three-year period evaluated for Denmark.

Background

The study compares manually engineered spectral-temporal features with annual TESSERA embeddings using Sentinel observations from 2020–2022. TESSERA exhibits less sensitivity to the selected input years than the engineered spectral-temporal representation, suggesting potential robustness when dense, high-quality local time series are unavailable.

However, the experiment does not disentangle the effects of cloud cover, observation density, phenological variability, and foundation-model pre-training. Because only three years are assessed, the authors state that broader robustness to year-specific data quality cannot be established. Resolving this issue requires evaluation across additional years and controlled variation in the quality and density of the underlying observations.