BetaEarth encoders 4 locations · real data
BetaEarth encodes Sentinel-2, Sentinel-1 and DEM separately, each into a 64-dimension vector per pixel. A separate step averages them. The average and the matching AlphaEarth tile are shown for four Major TOM locations, with one PCA colour basis fit on AlphaEarth across every tile.
inputs
encode
BetaEarth
encoder · DNN
embeddings
combine
Σ÷3
average
3 embeddings
Average vs AlphaEarth
⇄
BetaEarthAlphaEarth
Switch on at least one encoder
drag to compare · H × W × 64 · Loch Lomond
Data from Major TOM: lowest-cloud Sentinel-2 (L2A and L1C), Sentinel-1 RTC (VV and VH), and Copernicus DEM. Each sensor is encoded by the default model, segformer-film-robust. The average is the per-pixel mean of the selected embeddings, L2-normalised.