Combining hyperspectral and standard satellite images in one model

Hyperspectral satellites capture far more spectral detail than standard multispectral ones, but no foundation model has combined them effectively. SpectralEarth-FM uses a hierarchical transformer with spectral tokenization and cross-sensor fusion to handle all five data types simultaneously—hyperspectral, optical, thermal, and radar. The team curated a 40TB dataset linking hyperspectral images from three satellites (EnMAP, EMIT, DESIS) with co-located Sentinel and Landsat data, then pretrained on 25M patches using a contrastive objective. It tops benchmarks for both hyperspectral tasks and standard Earth observation metrics.