Research notes on super-resolution and land-use classification.
Land use and land cover
A framework for high-resolution land-use and land-cover classification using AlphaEarth embeddings
Existing land-cover products are coarse and years out of date. LUCERA takes CORINE's class definitions into the AlphaEarth embedding space, then densifies and extrapolates them — no spectral signatures, no temporal composites, one model for every region.
Super-resolution
S2DR3: effective 12-band 10× single image super-resolution for Sentinel-2
S2DR3 upscales every band of a single Sentinel-2 L2A scene from 10, 20 and 60 m/px to 1 m/px. This article documents its spectral fidelity, the limits of its spatial reconstruction, and the metrics used to evaluate both.
Super-resolutionSuperseded
S2DR2: effective 12-band 10× single image super-resolution for Sentinel-2
The first public case for treating Sentinel-2 super-resolution as a measurement rather than a decoration: what 1 m/px means when the sensor is 10 m/px, why the reconstructed detail is not recoverable directly from the source scene, and where the model fails.
Each article is reproduced here in full from its original post, which stays the canonical version. The long per-dataset evaluation galleries were left on Medium; every figure the text refers to is on this site. Written by the GammaEarth team.
